ARTIFICIAL INTELLIGENCE – KNOWLEDGE
Understand AI – beforehand and in action.
Solid knowledge for management and IT in businesses, organisations and institutions – so your move into professional AI rests on substance, and you stay in control of how it runs.
⌕Search the knowledge base — find the right articleOUR TOPICS:
Fundamentals & Orientation
Terms explained without jargon.
Retrieval-Augmented Generation (RAG): how AI answers from your own documents
RAG couples a language model to a searchable knowledge base. Answers come with source attribution, not invented from training.
EMBEDDINGS · AI CONCEPTEmbeddings and vectors: how language becomes mathematics
Embeddings are numerical representations of text, image, or audio. They are the foundation of every semantic search and every RAG system.
PROMPTING · AI CONCEPTPrompt engineering: foundations, patterns, anti-patterns
System prompt, few-shot, structured outputs, refusal patterns. What still matters in May 2026 – and what models now handle internally.
LLM GATEWAY · AI CONCEPTWhat is an LLM gateway? Purpose, components, market status May 2026
An LLM gateway is a central proxy for language-model calls. It bundles routing, auth, rate limits, fallback, observability and cost tracking.
VECTOR INDEX · AI CONCEPTWhat is a vector index? HNSW, IVF, ScaNN and quantisation May 2026
A vector index is the data structure inside a vector DB that finds similar embeddings fast. Trade-off between recall, latency and memory.
FINE-TUNING vs RAG · AI CONCEPTFine-Tuning vs RAG: which approach fits when? Status May 2026
Fine-tuning changes model behaviour permanently, RAG injects fresh knowledge. PEFT/LoRA makes FT affordable; RAG stays standard in compliance.
AI AGENT · AI CONCEPTWhat is an AI agent? ReAct, tool use and production patterns May 2026
An AI agent is an LLM system that calls tools itself, plans, and acts in multiple steps. Patterns May 2026: tool use, ReAct, LangGraph.
MCP · AI CONCEPTWhat is MCP (Model Context Protocol)? Tool-use standard May 2026
MCP is an open standard for LLM access to tools, data and servers. Anthropic spec Nov 2024, adopted by OpenAI/Google/Microsoft as of May 2026.
QUANTISATION · AI CONCEPTWhat is quantisation? Compressing model weights without quality loss
Quantisation stores model weights in fewer bits. Q4_K_M shrinks Llama-70B from 140 GB to 42 GB at under 2% quality loss.
TRANSFORMER · AI CONCEPTWhat is the transformer architecture? Basics, variants, market status May 2026
The transformer architecture is the technical foundation of all modern language models. Explained: self-attention, encoder-decoder, multi-head, MoE trend May 2026.
CONTEXT WINDOW · AI CONCEPTWhat is the context window? Token limit, cost, recall curve May 2026
The context window is the maximum token amount per LLM request (input + output). In May 2026: 128k-2M depending on provider, with a clear recall drop from the middle.
TOKEN · AI CONCEPTWhat is a token? Tokenisers, cost, DE-vs-EN May 2026
A token is a word fragment – the smallest billing unit of an LLM. Explained: BPE, SentencePiece, Tiktoken, German overhead, May 2026 price examples.
ATTENTION · AI CONCEPTWhat is the attention mechanism? Query, key, value explained May 2026
Attention is the heart of modern language models: every position of a sequence may attend to every other. Explained: Q/K/V, self vs cross, multi-head.
TEMPERATURE / TOP-P · AI CONCEPTWhat are temperature and top-p? LLM sampling parameters explained May 2026
Temperature, top-p and top-k control how deterministic or creative an LLM answers. Rules of thumb May 2026: 0-0.3 for facts, 0.7 for standard, 1.0+ for creative.
SYSTEM PROMPT · AI CONCEPTWhat is a system prompt? Role, security, best practices May 2026
A system prompt defines role, task and behaviour of an LLM before every user request. Explained: stages, prompt injection, Constitutional AI May 2026.
LLM BASICS · AI CONCEPTHow does an LLM work? Autocomplete on steroids, explained for SMEs May 2026
A language model predicts the most likely next word fragment. Explained in five stations: tokenisation, embedding, transformer, logits, sampling – without maths.
PRETRAINING · AI CONCEPTWhat is pretraining? How an LLM learns its base capability May 2026
Pretraining is the self-learning phase in which a language model absorbs language, grammar and factual knowledge from 10-15 trillion tokens of text. Explained: data sources, cutoff, cost.
RLHF · AI CONCEPTWhat is RLHF? Reinforcement Learning from Human Feedback explained May 2026
RLHF turns a raw base model into a helpful assistant. Three phases: SFT, reward model, PPO. Plus comparison with DPO, Constitutional AI and RLAIF May 2026.
MIXTURE OF EXPERTS · AI CONCEPTWhat is Mixture of Experts (MoE)? Sparse models explained May 2026
MoE models activate only a fraction of their parameters per token – comparable quality at 5x less compute. May 2026: Llama 4 Maverick, the current DeepSeek-V generation, Mixtral.
MULTIMODAL · AI CONCEPTWhat is multimodal AI? Image, audio, video plus text May 2026
Multimodal models process not only text but also image, audio and video. May 2026: GPT-4o, Gemini 2.5 Pro, the current top Claude model, Llama 4. Use cases for document recognition and damage photos.
REASONING · AI CONCEPTWhat is a reasoning model? o3, Claude Extended Thinking, the current DeepSeek-R generation May 2026
Reasoning models think internally in chain-of-thought before answering. More tokens for thinking = better answers in maths, code, logic. Costs 5-15x more than regular models.
TOOL USE · AI CONCEPTWhat is tool use and function calling? LLMs invoking external APIs May 2026
Tool use turns a language model into an agent: the model structurally invokes external APIs – calculator, database, web search. May 2026 standard with MCP protocol.
OWN MODEL · AI CONCEPTHow to train your own AI model? Fine-tuning, LoRA, QLoRA May 2026
Fine-tuning vs from-scratch vs RAG: for 95% of SMEs fine-tuning with LoRA/QLoRA is the right path. Hardware needs, cost USD 5-50k, May 2026 tools.
GLOSSARY · AI TERMSAI glossary: 80+ key terms for SMEs and fiduciary firms
Glossary with 80+ AI terms from A to Z. Short definitions, links to deeper pages. As of May 2026.
FAQ · AI FOR SMESAI FAQ for SMEs: 35 frequent questions on practice, law and cost
Answers to the most frequent AI questions from Swiss SMEs: cost, data protection, tools, hallucinations, FINMA, EU AI Act. As of May 2026.
Basics · EconomicsWhat does AI really cost in a fiduciary firm? Total cost and payback by mandate size
Cost types, total cost of ownership and payback logic for AI in fiduciary firms – with calculation method, not promises.
Application by Domain
By domain: sales, HR, accounting …
WhatsApp & Telegram bot: AI answering on the channels your clients actually use
Working bot on WhatsApp Business or Telegram, tested and live. Appointment booking, FAQ, lead qualification, connected to CRM or RAG knowledge base. Flat fee CHF 1,800.
VOICE · SERVICEVoice agent on the phone: AI that calls and is called
Phone agent with Whisper STT, LLM, and ElevenLabs/Cartesia TTS. Call answering, appointment booking, pre-qualification. Latency budget under 800 ms. Flat fee CHF 3,500.
PAYROLL TRIAGE · USE CASEAI triage in payroll: pre-sorting client questions on AHV, BVG, and withholding tax
AI pre-sorts incoming payroll queries, drafts answers from official guidelines, and hands the case with context to the case handler.
VAT PREPARATION · USE CASEAI-assisted VAT preparation: classifying receipts, suggesting input-tax codes, checking the net tax rate method
AI classifies receipts by VAT code, proposes input-tax deduction, and flags net-tax-rate branches. The fiduciary checks, corrects and books.
CLIENT TRIAGE · USE CASEAI triage for client queries: turning WhatsApp, email and phone into structured cases
AI classifies client queries, finds answer building blocks in the internal knowledge base, and prepares a draft for the case handler.
RECEIPT OCR · USE CASEAI receipt recognition for Swiss documents: structured capture of QR-bills, receipts and PDF invoices
OCR extracts structured fields from QR-bills, restaurant receipts and PDF invoices, validates IBAN and UID, and hands off to the ERP.
YEAR-END QA · USE CASEAI-supported quality assurance for the annual financial statement
Detect anomalies in journal entries before the audit: Benford test, balance comparison, accruals. The human decides, the AI flags.
TAX DRAFT · USE CASEAI draft for tax optimisation of legal entities
Draft variants for dividend vs salary, participation deduction, loss offset. Tax advisor finalises and signs. NOT an end product for the client.
COLLECTIONS · USE CASEAI-supported collections without damaging the client relationship
Tiered reminders 1-2-3, deferral request triage, personalised letters from client history. Integration with Bexio, Abacus, Banana. The fiduciary decides before every dispatch.
AML SCREENING · USE CASEAI-supported AML/KYC screening: sanctions lists, PEP, beneficial owners
Screen onboarding data against SECO/OFAC/EU lists, build a risk score, prepare an SRO report draft. NEVER report automatically – the human decides.
FIDUCIARY · INDUSTRY HUBAI for Swiss fiduciary offices: a practical guide
Where AI today actually unburdens Swiss fiduciary work in payroll, VAT, AML and collections – without touching professional secrecy.
LAW FIRM · INDUSTRY HUBAI for Swiss law firms: what works, what does not, and what the SAV guidance says
How Swiss law firms use AI in contract analysis, case research and client communication – without breaching attorney privilege under Art. 13 LFLP.
INSURANCE · INDUSTRY HUBAI for Swiss insurance: governance, claims and underwriting
How Swiss insurers and brokers deploy AI across claims, underwriting and client queries – within FINMA Supervisory Notice 08/2024.
MEDICAL PRACTICES · INDUSTRYAI for medical practices: dictation, correspondence, triage – what is legally allowed and what is not
Dictation, KVG correspondence and triage relieve Swiss medical offices – patient data is highly sensitive and falls under EU AI Act high-risk rules.
ARCHITECTURE · INDUSTRYAI for architects and planners: BIM, building applications, tenders – where language models actually save time
Building applications, tender reading, contracts and research eat hours in every SIA office. Language models cut measurable time – structural engineering stays with the engineer.
TRADES · INDUSTRYAI for Swiss trades: quotes, dunning, WhatsApp triage – where ROI is fastest
Quotes in 10 minutes instead of 60, automated dunning, incoming WhatsApp messages pre-sorted – Swiss trades extract AI value fastest. Setup pays back in 3 to 6 months.
REAL ESTATE · INDUSTRYAI for Swiss real-estate brokers and property managers: listings, leases, applicant triage – and AML obligations
Listings, applicant pre-qualification and contracts are routine in every Swiss brokerage. AI speeds them up – but brokerages are AML-bound and applicant scoring sits in EU AI Act high-risk territory.
E-COMMERCE · INDUSTRYAI for Swiss e-commerce: product copy, customer triage, recommendation engines – and revDSG for tracking
Multilingual product copy, FAQ bots and recommendation engines win margin back. Profiling scores sit under EU AI Act watch, tracking needs revDSG-compliant consent.
EMAIL TRIAGE · USE CASEEmail triage automation: classify inbound flood, assign to client, prepare draft
IMAP watcher reads every mail, EU LLM classifies (client/invoice/query/newsletter/spam), RAG attaches client context, draft lands with the case handler. Dispatch only by hand.
QUOTES · USE CASEAI-assisted quote generation: 2 to 4 hours of work in 20 minutes
From inquiry -> RAG over past quotes (price ladders, standard clauses) -> clean draft in Bexio/Klara format. Case handler reviews and sends manually.
CH MULTILINGUALISM · COMPLIANCEAI and Swiss multilingualism: LLMs for German, French, Italian and Romansh
Swiss language split (2025): German 62.8%, French 22.9%, Italian 8.2%, Romansh 0.5%. Which models handle the four national languages plus Swiss German – as of May 2026.
SWISSDEC ELM · COMPLIANCESwissdec ELM 5.0 and electronic wage statement: certified payroll software, interfaces, AI plausibility checks
Swissdec ELM standard 5.0 is in production as of May 2026. Certified payroll software, interfaces to AHV, SUVA, health funds, pension funds, tax. AI for wage-statement plausibility checks.
VAT · COMPLIANCEVAT and AI automation in Switzerland: receipt recognition, VAT code suggestion, input-tax optimisation, ESTV XML
CH VAT rates since 1 Jan 2024: standard 8.1%, reduced 2.6%, accommodation 3.8%. AI for receipt OCR, VAT code suggestion, input-tax optimisation. ESTV quarterly filing via SuisseTax/ePortal.
BANKING · INDUSTRY HUBAI for Swiss banks under FINMA supervision: governance, credit scoring, AML and client triage
How Swiss banks use AI in credit scoring, AML/KYC, sentiment analysis and client mail triage – within FINMA 08/2024, Banking Act Art. 47 and the revised FADP.
ASSET MANAGEMENT · INDUSTRY HUBAI for Swiss independent asset managers: suitability, reporting and market synthesis
How FinIA-licensed Swiss asset managers use AI in risk profiling, reporting automation and market news synthesis – within FinSA, FINMA 08/2024 and the revised FADP.
NOTARIES · INDUSTRY HUBAI for Swiss notaries: precedent research, civil-code templates and notarisation preparation
How Swiss notaries use AI for case-law research, notarisation preparation and Civil Code estate templates – with anonymisation duty and without delegating the notarisation act itself.
LITIGATION · INDUSTRY HUBAI for litigation firms: eDiscovery, case analysis, brief preparation and Swiss case law
How Swiss litigation firms and disputes boutiques use AI in eDiscovery, case analysis, brief preparation and case-law research – with BGFA and Art. 321 SCC as guardrails.
OCCUPATIONAL HEALTH & SOCIAL INSURERS · INDUSTRY HUBAI for Swiss social, health and accident insurers and occupational health services
How Swiss health and accident insurers and occupational health services use AI in claims handling, disability-claim plausibility and medical-officer document analysis – with especially sensitive personal data (revised FADP Art. 5).
MEP & BUILDING PLANNING · INDUSTRY HUBAI for building planning and MEP/HVAC: SIA standards search, schema generation, energy modelling and BIM
How Swiss MEP/HVAC and building planners use AI for SIA standard search, HVAC schema generation, energy modelling and BIM enrichment – within SIA standards, MuKEn 2014 and Minergie.
PHARMA & MEDTECH · INDUSTRY HUBAI for pharma and medtech: regulatory RAG, adverse-event triage and AI as medical device
How Swiss pharma and medtech companies use AI in regulatory research, adverse-event triage and literature synthesis – within HMG, Swissmedic, ISO 13485, MDR EU 2017/745 and the EU AI Act.
Application · Document ProcessingReading and Booking QR-Bills & eBill/Peppol Automatically with AI
How fiduciaries read QR-bills, eBill and Peppol documents in a structured way and book them semi-automatically with AI-assisted coding – human in the loop.
ApplicationAutomated Account Assignment & the SME Chart of Accounts: AI Booking Suggestions with Human-in-the-Loop
How AI generates booking suggestions on the Swiss SME chart of accounts, with a feedback loop and human approval – and where the error sources lie.
Application · Payroll & Social InsuranceAI in Social Insurance: Automating AHV/IV/EO, BVG, Withholding Tax & Swissdec ELM
Where AI usefully supports Swiss payroll (AHV, ALV, BVG, UVG, withholding tax, Swissdec ELM) – and where calculation must stay deterministic.
Application · Tax & FiduciaryAI-Assisted Tax Filing & Document Capture: Tax Software, Dr. Tax and Your Own RAG Workflow
How Swiss tax software, OCR capture and a private RAG workflow work together – including limits, responsibility and data protection.
Law, Compliance, GDPR
AI Act, liability, co-determination.
revDSG / revFADP and AI: what the revised Swiss Data Protection Act means for LLM use
The revised Swiss FADP has been in force since 1 Sep 2023. Anyone running LLMs on personal data must observe Art. 4, 5, 6, 7, 16-18, 19 and 21.
EU AI ACT · COMPLIANCEEU AI Act 2026: high-risk duties from 2 August 2026 – what Swiss providers must do now
Regulation (EU) 2024/1689 phases in. Prohibitions since 2 Feb 2025, GPAI since 2 Aug 2025, high-risk from 2 Aug 2026 – also for Swiss providers serving EU clients.
AMLA REVISION · COMPLIANCEAMLA revision 2026: extension to fiduciary advisory and FATF Recommendation 16
The Federal Council adopted the AMLA revision in 2024. Extension to advisory fiduciary work, new SRO duties, FATF Recommendation 16 – entry into force H2 2026.
ART. 957a CO · COMPLIANCEArt. 957a CO and AI bookings: audit trail, GeBüV, and 10-year retention
Art. 957a CO demands reliability, completeness, traceability. AI-assisted bookings need an audit trail with model version, prompt, source, and human review.
ART. 321 SCC · COMPLIANCEProfessional secrecy (Art. 321 SCC) and AI use: what lawyers, notaries, physicians and auditors must observe
Art. 321 SCC forbids disclosure of professional secrets. Cloud LLM use can be such a disclosure. When AI tools remain admissible.
ISO 42001 · COMPLIANCEISO/IEC 42001: the international standard for AI management systems
ISO/IEC 42001:2023 is the first international AIMS standard. Who needs certification, what it costs, which Annex A controls are mandatory.
TIA · COMPLIANCEThird-country transfer and Transfer Impact Assessment (TIA): Swiss data in US and PRC cloud LLMs
Art. 16/17 revFADP, EU-US DPF with Swiss annex, EDÖB TIA module. When SCC suffice, when BCR are needed, what to assess at OpenAI/Anthropic.
FINMA · COMPLIANCEFINMA awareness: AI governance for banks, insurers and asset-managing fiduciaries
FINMA Supervisory Notice 08/2024, Circular 2023/01 on operational risks. Who is supervised, what FINMA requires, what is best practice for the non-supervised.
AUDIT TRAIL · AI CONCEPTAI audit trail design: what to log so an AI answer stays audit-ready
Which fields must be stored per LLM call so you stay clean under Art. 957a CO and professional secrecy – and with which tools.
GDPR · COMPLIANCEGDPR and LLMs: when the EU General Data Protection Regulation applies directly to Swiss companies
Under Art. 3(2) the GDPR reaches Swiss firms whenever EU customers are involved. LLM use touches practically every risk article.
DPIA · COMPLIANCEDPIA for AI systems: Data Protection Impact Assessment under revDSG Art. 22 and GDPR Art. 35
A DPIA is mandatory for AI with profiling, high-risk data or automated decisions. Methodology: CNIL PIA + EDPB template + AI Act overlay.
AI LIABILITY · COMPLIANCEWho is liable for AI errors? Swiss CO + EU Product Liability Directive 2024/2853 after the AI Liability Directive withdrawal
Swiss liability runs via CO Art. 41/55/97. The EU AI Liability Directive was withdrawn in Feb 2025. PLD 2024/2853 in force since 9 Dec 2024 – software incl. AI is covered as a product.
EU AI ACT SME · COMPLIANCEEU AI Act for SMEs: deadlines and mandatory check by 2 Aug 2026
High-risk duties from 2 August 2026. How a Swiss SME deployer works through the AIA timeline concretely – inventory, classification, training.
revDSG · FIDUCIARY COMPLIANCErevDSG and AI in the fiduciary office: DPA, data categories, FDPIC position, GwG practice
Which DPA with OpenAI/Anthropic/Mistral, which client data can go where, FDPIC 2026 position on LLMs, sensitivity-based routing. Not legal advice.
ANONYMISATION · AI CONCEPTAnonymisation and pseudonymisation: Presidio, Privacera, k-anonymity, differential privacy
Tools and techniques as of May 2026 for Swiss-DSG-compliant PII removal before LLM processing: Microsoft Presidio, Privacera, Anonymizer, k-anonymity and differential privacy compared.
FDPIC · COMPLIANCEFDPIC notification duty and AI guidance: 72-hour rule under FADP Art. 24 and current recommendations 2024-2026
FDPIC 72-hour breach duty from awareness (FADP Art. 24). AI opinions 2024-2026. Addresses, forms and escalation path for Swiss fiduciary, law and SME operations.
FINMA SN 08/2024 · COMPLIANCEFINMA Supervisory Notice 08/2024 on AI: four pillars governance, accountability, robustness, explainability
Supervisory Notice 08/2024 of 18 Dec 2024. Fiduciary relevance: only AMLA-supervised mandates. Pure accounting fiduciaries are not directly FINMA-supervised – but best practice is recommended.
Law & ComplianceMay I use ChatGPT as a Swiss fiduciary? Data protection, DPA & business version (revFADP + possibly Art. 321 SCC)
ChatGPT in the fiduciary office: when revFADP (and possibly secrecy under Art. 321 SCC) permit use - and why only Business/Enterprise with a DPA qualifies.
Law & ComplianceEU AI Act for Swiss SMEs 2026: Obligations, Risk Classes, Roadmap Checklist
When EU AI Regulation 2024/1689 reaches Swiss firms, which risk classes apply, and which 2026 deadlines are due.
Data Protection & Data SovereigntyUS CLOUD Act vs. Swiss Data Location: Why CH/EU Hosting Matters for Client Data
The US CLOUD Act lets US authorities access data from US providers – even in CH/EU data centres. What that means for client data and data location.
Law & ComplianceAI in Statutory Audit: What Is Permissible, What Compromises Independence & Audit Quality
Where AI is admissible as a tool in limited and ordinary audits in Switzerland, and where judgment and responsibility must remain with the auditor.
Tools & Providers
Providers compared, neutral.
LiteLLM: one gateway for 100+ LLM providers behind a single API
LiteLLM is an open-source proxy that bundles OpenAI, Anthropic, Mistral, local models and more behind a single OpenAI-compatible API.
QDRANT · TECHQdrant: production vector database for RAG and semantic search
Qdrant is an open-source vector database written in Rust. CPU-only, filter-capable, with payload indexes and stable metrics under multi-tenant load.
N8N · TECHn8n: workflow automation with 600+ integrations, self-hostable under EU law
n8n is a fair-code-licensed workflow platform with 600+ connectors, AI nodes, and full self-hosting on Docker.
OLLAMA · TECHOllama: local LLMs on your own hardware – where it works and where it does not
Ollama is a local runtime for open-source LLMs. Strong for privacy demos and CPU classification, slow for 70B models without GPU.
HETZNER · TECHHetzner as EU hosting for Swiss fiduciaries and SMEs: data centres, contracts, cost
Hetzner Online runs data centres in Falkenstein and Helsinki with ISO 27001, a data processing agreement, and prices from EUR 4/month cloud to EUR 39/month dedicated.
DOCKER · TECH STACKDocker orchestration for SMEs: docker-compose without Kubernetes overkill
Docker + docker-compose is enough for single-host setups up to 50 containers. Digest pinning, healthchecks, restart policies, resource limits, non-root.
NGINX · TECH STACKNginx as reverse proxy: SSL, rate limits, and security headers for containerised apps
Nginx 1.28 fronts Docker backends as an edge layer. Certbot SSL, gzip/brotli, limit_req zones, Cloudflare IP allowlist, HSTS/CSP, websocket upgrade.
CLOUDFLARE · TECH STACKCloudflare as DNS, reverse proxy, and WAF: SSL modes, cache rules, origin certificates
Cloudflare provides DNS, WAF, and DDoS protection for 9+ Fairlane domains. Proxied mode, Full-strict SSL, free 15-year origin cert, Cache Rules over Page Rules.
GRAFANA · TECH STACKGrafana, Prometheus, Loki: monitoring stack for container apps and LLM workflows
Grafana 11 plus Prometheus 3 plus Loki 3 as a self-hosted monitoring stack. Metrics via node-exporter and /metrics, logs via Promtail, alerts via Alertmanager to Telegram.
BREVO · TECH STACKBrevo SMTP: transactional email from EU hosting for SME apps
Brevo (formerly Sendinblue) provides SMTP relay smtp-relay.brevo.com:587 from FR data centres. DKIM/SPF/DMARC, revDSG-aligned, free 300 mails/day, paid from EUR 19/month.
OPENAI · LLM PROVIDEROpenAI GPT models from a Swiss fiduciary perspective: residency, pricing, compliance
OpenAI is the most widespread model ecosystem. For Swiss fiduciaries, the decisive points are: clarify data residency via Azure OpenAI and DPA posture before client data flows.
ANTHROPIC · LLM PROVIDERAnthropic Claude from a Swiss fiduciary perspective: residency, pricing, compliance
Anthropic Claude offers long context windows and cautious answers. For Swiss fiduciaries, the clean path leads through AWS Bedrock Frankfurt.
MISTRAL · LLM PROVIDERMistral AI from a Swiss fiduciary perspective: EU residency, pricing, sovereignty
Mistral is the only frontier lab headquartered in the EU. For Swiss fiduciaries the shortest path to GDPR-compliant model access without U.S. CLOUD Act residual risk.
COHERE · LLM PROVIDERCohere from a Swiss fiduciary perspective: RAG specialist with BYOC option
Cohere is not a chatbot provider but the RAG specialist. Rerank 3 is the industry standard, embed-multilingual-v3 strong for German.
GOOGLE GEMINI · LLM PROVIDERGoogle Gemini in Swiss practice: Vertex AI, Zurich region and data flow
Gemini 2.5 Pro/Flash/Flash-Lite via Vertex AI in europe-west6 (Zurich) or europe-west3 (Frankfurt). What that means for revDSG, price and model choice.
META LLAMA · LLM PROVIDERMeta Llama in Swiss practice: open-weight model, self-host or provider
Llama 4 Scout/Maverick and Llama 3.3 70B as the open-weight option. Licence, hardware needs, prices at Groq/Together/Fireworks and self-host reality.
DEEPSEEK · LLM PROVIDERDeepSeek in Swiss practice: PRC provider, self-host option and revDSG reality
DeepSeek V3.x and R1 are extremely cheap and technically strong – but data flow goes to China. Not recommended for client data via API. Self-host via HF weights as alternative.
SELF-HOSTED OLLAMA · LLM PROVIDERSelf-hosted Ollama as an LLM provider: when does it replace OpenAI, Anthropic or Gemini?
Self-hosted Ollama on a Hetzner GPU or office server: pays off from 2-5M tokens/month, replaces cloud LLMs for revDSG-sensitive workloads, has clear quality limits.
VECTOR DB · AI CONCEPTVector databases compared: Qdrant, Weaviate, Milvus, Pinecone, Chroma, pgvector
Six serious options, three architectural axes, one concrete recommendation per use-case. As of May 2026.
ROUTING · AI CONCEPTMulti-LLM routing: which model when, for how much
Routing rules by sensitivity, cost, latency, and quality. Fallback chain, semantic caching, cost observability. May 2026 pricing.
SELF-HOSTED VS. CLOUD · AI CONCEPTSelf-hosted vs. cloud LLM: a decision framework for SMEs and fiduciaries
When does running your own language model on your own hardware pay off, and when is the cloud the right choice? Total cost of ownership, latency, data protection.
VECTOR DATABASES · COMPARISONVector databases compared: 10 options for RAG, search, and recommendation
Qdrant, Weaviate, Milvus, Chroma, pgvector, Pinecone, Redis, Vespa, LanceDB and Elasticsearch in a neutral comparison, with hosting, license, and EU aspects.
LLM GATEWAYS · COMPARISONLLM gateways compared: 10 options for routing, audit, and cost control
LiteLLM, OpenRouter, Portkey, Kong, Cloudflare, Helicone, TrueFoundry, Martian, Bifrost, and Apache APISIX in a neutral comparison.
WORKFLOW AUTOMATION · COMPARISONWorkflow automation compared: 10 platforms for SMEs and fiduciaries
n8n, Make, Zapier, Pipedream, Activepieces, Windmill, Temporal, Airflow, Node-RED, and Huginn in a neutral comparison.
EMBEDDINGS · TOOL COMPARISONEmbedding models compared: BGE-M3, E5, OpenAI, Cohere, Voyage, Jina, Mistral, Nomic, mxbai, Gecko
Ten serious embedding models, four selection axes, one concrete recommendation per use-case. As of May 2026.
RERANKER · TOOL COMPARISONRerankers compared: Cohere, BGE, Jina, Voyage, ColBERT, mxbai, Mistral, sentence-transformers, RankGPT, FlashRank
Ten reranker options, four selection axes, +15-30% recall lift for RAG pipelines. As of May 2026.
LLM PROVIDERS · TOOL COMPARISONLLM providers compared: OpenAI, Anthropic, Mistral, Google, Cohere, Meta, DeepSeek, xAI, AWS Bedrock, Azure
Ten LLM providers in direct comparison, seven selection axes, one concrete recommendation per use-case. As of May 2026.
LOCAL LLM RUNTIMES - COMPARISONLocal LLM runtimes compared: Ollama, vLLM, llama.cpp, LM Studio, LocalAI, TGI, GPT4All, KoboldCpp, Jan, OpenLLM
Ten serious runtimes for locally operated language models, from hobby desktop to production GPU serving. Decision matrix as of May 2026.
OPEN-WEIGHT MODELS - COMPARISONOpen-weight models compared: Llama 3.3/4, Mistral, DeepSeek, Qwen, Gemma, Phi-4, Command R, Falcon, GLM, Apertus
Eleven open-weight model families plus Switzerland's Apertus, available as of May 2026. Licence, VRAM need, multilingual capability, practice recommendation.
SOVEREIGN HOSTING - COMPARISONSovereign hosting compared: Hetzner, Infomaniak, Exoscale, OVHcloud, Scaleway, Swisscom, Safe Swiss Cloud, netcup, Contabo, on-prem
Ten hosting options for AI workloads in Switzerland and the EU. Data residency, CLOUD Act, GPU availability, FINMA suitability. As of May 2026.
GPU CLOUD · TOOL COMPARISONGPU cloud providers compared: RunPod, Vast.ai, Lambda, CoreWeave, Paperspace, Exoscale, Hetzner, Together, Replicate, Modal
Ten serious options for GPU hours, from auction marketplace to enterprise premium. As of May 2026 with current prices per H100, A100, and RTX 4090.
CONTAINER DEPLOY · TOOL COMPARISONContainer deployment compared: Docker, Podman, Kubernetes, Swarm, Coolify, Dokku, CapRover, Nomad, Portainer, Railway/Render
Ten ways to run containers on a server – from a single Docker daemon to a Kubernetes cluster. With clear SME recommendations as of May 2026.
REVERSE PROXY · TOOL COMPARISONReverse proxies compared: nginx, Caddy, Traefik, HAProxy, Apache, Envoy, nginx Proxy Manager, Cloudflare Tunnel
Eight serious reverse-proxy options in direct comparison. With clear per-use-case recommendation, as of May 2026.
MONITORING / TOOL COMPARISONMonitoring & observability compared: Grafana, Loki, Uptime Kuma, Netdata, Zabbix, Datadog, Sentry, ELK, VictoriaMetrics, SigNoz
Ten serious tools for metrics, logs, traces, and error tracking. Seven decision axes, one concrete recommendation per SME scenario. As of May 2026.
LLM OBSERVABILITY / TOOL COMPARISONLLM observability compared: Langfuse, Helicone, LangSmith, Phoenix, Lunary, Portkey, OpenLLMetry, Traceloop, HoneyHive, W&B Weave
Ten specialised tools for tracing, cost tracking, prompt versioning, and evaluation of LLM pipelines. Seven decision axes, one concrete recommendation per scenario. As of May 2026.
DNS / CDN / WAF / TOOL COMPARISONDNS, CDN and WAF compared: Cloudflare, Bunny.net, Fastly, AWS CloudFront, Akamai, Gcore, Infomaniak, DNSimple
Eight serious providers for DNS, CDN, and web application firewall. Seven decision axes, one concrete recommendation per use case. As of May 2026.
SMTP COMPARISON · TOOL COMPARISONTransactional email services compared: Brevo, Postmark, Mailgun, SES, SendGrid, Resend, Mailjet, Infomaniak, SMTP2GO
Nine serious options for sending APIs. Pricing, deliverability, data residency for Swiss SMEs and fiduciary offices. As of May 2026.
DB COMPARISON · TOOL COMPARISONDatabases compared: PostgreSQL, MySQL/MariaDB, SQLite, MongoDB, Redis, ClickHouse, CockroachDB, SurrealDB, DuckDB, Supabase
Ten serious databases for SME setups. Relational vs. NoSQL vs. columnar, self-host vs. cloud, with clear per-use-case recommendations. As of May 2026.
SECURITY COMPARISON · TOOL COMPARISONSecurity hardening tools compared: CrowdSec, Fail2ban, Wazuh, UFW, Vault, Authentik, WireGuard, Lynis, rkhunter, ClamAV
Ten open-source tools for server hardening, intrusion detection, SSO and secrets management at Swiss SMEs. Clear per-layer recommendations. As of May 2026.
GIT & CI/CD · TOOL COMPARISONGit hosting and CI/CD compared: Gitea, GitLab, Forgejo, GitHub, Drone, Woodpecker, Jenkins, Act, ArgoCD
Nine serious tools for code hosting, pipelines, and GitOps. Self-host vs cloud, learning curve, SME fit. As of May 2026.
BOT & VOICE · TOOL COMPARISONBuilding blocks for chat and voice bots compared: Whisper, Deepgram, ElevenLabs, Piper, Twilio, Vapi, Retell, WhatsApp, Rasa, Botpress
Ten building blocks for voice and chat bots. STT, TTS, telephony, voice-AI platforms, and chatbot frameworks compared directly. As of May 2026.
RAG FRAMEWORKS · TOOL COMPARISONRAG frameworks compared: LangChain, LlamaIndex, Haystack, DSPy, Semantic Kernel, txtai, RAGFlow, Verba, Flowise, Langflow
Ten serious frameworks for RAG pipelines. Code-first, visual builders, and academic approaches compared directly. As of May 2026.
N8N · MAKE · ZAPIER · DUELn8n vs Make vs Zapier – which fits Swiss SMEs?
Three workflow-automation platforms head-to-head: self-hosting, price per 1000 runs, EU residency, AI nodes. Decision guide for fiduciary, legal, and SMEs.
OPENAI · ANTHROPIC · MISTRAL · DUELOpenAI vs Anthropic vs Mistral – which LLM provider in 2026?
Three LLM providers head-to-head: GPT-Modelle (4o, 4.1, o-Reihe), the current Claude model, Mistral Large 2/Small 3.1. Pricing, EU region, DACH language, lock-in. Decision guide May 2026.
QDRANT · WEAVIATE · PGVECTOR · DUELQdrant vs Weaviate vs pgvector – which vector DB for SME RAG?
Three vector databases head-to-head: self-host effort, scaling limit, Postgres integration, EU hosting. Decision guide for RAG pipelines in Swiss SMEs.
LITELLM · OPENROUTER · PORTKEY · DUELLiteLLM vs OpenRouter vs Portkey – which LLM gateway?
Three LLM gateways head-to-head: self-host vs cloud, number of providers, compliance routing, EU hosting. Decision guide for multi-LLM setups in Swiss SMEs.
HETZNER · INFOMANIAK · EXOSCALE · DUELHetzner vs Infomaniak vs Exoscale – where does a Swiss fiduciary host its AI?
Three hosters head-to-head: CH data residency, monthly price, GPU availability, compliance profile. Decision guide for AI workloads in Swiss SMEs.
OLLAMA vs vLLM vs LLAMA.CPP - DUELOllama vs vLLM vs llama.cpp - which local LLM server?
Three open-source runtimes for local language models. Ollama for onboarding, vLLM for production throughput, llama.cpp as the portable foundation - decision matrix as of May 2026.
LANGCHAIN vs LLAMAINDEX vs HAYSTACK - DUELLangChain vs LlamaIndex vs Haystack - which RAG stack in 2026?
Three OSS frameworks for retrieval-augmented generation. LangChain broad but noisy, LlamaIndex clean for RAG, Haystack enterprise-grade - decision matrix as of May 2026.
LANGFUSE vs HELICONE vs LANGSMITH - DUELLangfuse vs Helicone vs LangSmith - which LLM tracing tool?
Three LLM observability platforms. Langfuse as the EU OSS standard, Helicone as the 5-minute proxy, LangSmith for LangChain stacks - decision matrix as of May 2026.
RUNPOD vs VAST.AI vs HETZNER GPU - DUELRunPod vs Vast.ai vs Hetzner GPU - where to train and infer in 2026?
Three GPU cloud models. RunPod as a serious on-demand platform, Vast.ai as a spot marketplace, Hetzner as EU-reserved hosting - price and profile comparison as of May 2026.
CLOUDFLARE vs BUNNY.NET vs FASTLY - DUELCloudflare vs Bunny.net vs Fastly - which edge platform for SMEs?
Three CDN+DNS+WAF providers compared. Cloudflare with a free tier and the largest network, Bunny.net EU-focused and very cheap, Fastly enterprise-only - decision as of May 2026.
JSON & STRUCTURED OUTPUT · AI CONCEPTOutput formatting and JSON mode: function calling, Pydantic, Instructor, Outlines
Enforce structured LLM outputs: JSON mode, function calling, Pydantic parsing, Instructor library, Outlines (local) and constrained decoding for GPT-4.1, the current top Claude model, Mistral.
HYBRID SEARCH · AI CONCEPTHybrid search: BM25 plus vectors with reciprocal rank fusion in Elasticsearch, Qdrant, OpenSearch
Why pure vector search often fails in fiduciary contexts and how BM25 plus vectors with RRF or weighted sum measurably improves retrieval quality. Tools: Elasticsearch, Qdrant hybrid, OpenSearch.
QUERY EXPANSION · AI CONCEPTQuery expansion and rewriting: HyDE, decomposition, multi-query, step-back prompting
How to rewrite short user questions so RAG finds the right sources: HyDE, query decomposition, multi-query, step-back prompting. When it helps, when it does not.
METADATA · AI CONCEPTMetadata and filters in RAG: pre-filter vs post-filter, Qdrant payload index, pgvector WHERE
How structured metadata makes client, date, confidentiality, language and source filterable per chunk: pre- vs post-filter, Qdrant payload index, pgvector WHERE and time-aware retrieval.
FIDUCIARY SOFTWARE · COMPLIANCESwiss fiduciary software landscape: market share, API openness and AI maturity (May 2026)
Bexio, Abacus, Sage 50, KLARA, Run my Accounts, Topal, AbaWeb, Atlanto: who dominates in which SME segment and which tools are AI-ready.
SWISS CLOUD · COMPLIANCESovereign Swiss cloud hosting: Infomaniak, Exoscale, Swisscom, Safe Swiss Cloud, Hostpoint, Cloudsigma compared
Six Swiss cloud providers compared May 2026: region, ISO 27001, FINMA suitability, GPU availability, price. Practical guide to sovereign hosting in Switzerland.
APERTUS · COMPLIANCEApertus: the open Swiss AI model from ETH Zurich, EPFL and CSCS – status May 2026
Apertus 8B and 70B, Apache 2.0, from ETH/EPFL/CSCS. Released 2 Sep 2025, 15T tokens, 1000+ languages including Swiss German and Romansh. As of May 2026: production experience and Apertus 2 outlook.
WEAVIATE · TECHWeaviate: open-source vector database with GraphQL and multi-modal modules
Weaviate is an Apache-2.0 vector DB in Go with GraphQL and REST APIs. May 2026 v1.30+, EU region Frankfurt, native hybrid search, multi-modal modules for text and image.
MILVUS · TECHMilvus: cluster vector database for more than one billion vectors
Milvus is an Apache-2.0 vector DB with separated compute and storage layers. GPU acceleration, HNSW plus IVF plus DiskANN, for volumes from 100M vectors.
CHROMA · TECHChroma: the simplest vector database for prototypes and notebooks
Chroma is an Apache-2.0 vector DB with a DuckDB backend. Python-API first, productive in ten minutes, good for prototypes up to 1M vectors. Not for production scale.
PGVECTOR · TECHpgvector: vector search inside existing PostgreSQL with HNSW and IVFFlat
pgvector is the PostgreSQL extension for vector search. May 2026 v0.8+ with HNSW, IVFFlat, and binary quantisation. ACID, joins, one database for everything.
PINECONE · TECHPinecone: managed cloud vector DB without self-hosting
Pinecone is a proprietary cloud-only vector DB. EU region eu-west-1 available, serverless since 2024, zero ops. Vendor lock-in and USD pricing risk.
REDIS REDISEARCH · TECHRedis with RediSearch: vector index in the KV store for low latency
Redis with the RediSearch module ships an HNSW vector index since version 2.4. Useful when Redis already runs in the stack; fewer vector features than specialised DBs.
VESPA · TECHVespa: search engine with tensor ranking for complex hybrid pipelines
Vespa is an Apache-2.0 search engine in Java from the Yahoo ecosystem. Tensor ranking, structured plus vector plus full text in one query. Steep learning curve.
LANCEDB · TECHLanceDB: embedded vector DB in the columnar Lance format for local apps
LanceDB is an Apache-2.0 vector DB in Rust with columnar Lance format. Embedded in Python/JS, no server required, very fast. Good for desktop and small on-prem.
ELASTICSEARCH KNN · TECHElasticsearch with kNN: hybrid keyword and vector search in one query
Elasticsearch from version 8 offers native kNN vector search. May 2026 v9 with improved quantisation. Strong for hybrid search, Elastic License v2 / SSPL.
MEILISEARCH · TECHMeilisearch: fast search engine with AI search mode for hybrid search
Meilisearch is an MIT search engine in Rust. May 2026 v1.10+ with AI search mode (BM25+embedding), developer-friendly, good for search bars with AI snippets.
OPENROUTER · TECHOpenRouter: multi-model marketplace for fast LLM comparison
OpenRouter is a US cloud gateway with 200+ LLM models behind an OpenAI-compatible API, auto-fallback, and 5% markup on token pricing.
PORTKEY · TECHPortkey: enterprise LLM gateway with EU hosting and self-host tier
Portkey is a cloud gateway with self-host tier, EU region Frankfurt, 1,600+ LLMs, observability, caching, and guardrails for enterprise compliance.
KONG AI GATEWAY · TECHKong AI Gateway: Kubernetes-native API gateway with LLM plugins
Kong v3.8 extends the open-source API gateway with AI-Proxy, AI-Prompt-Guard, and semantic caching – self-host on Kubernetes or bare metal.
CLOUDFLARE AI GATEWAY · TECHCloudflare AI Gateway: edge proxy for OpenAI, Anthropic, Workers AI
Cloudflare AI Gateway runs on the Cloudflare edge, is free in the Workers plan, and bundles OpenAI, Anthropic, Mistral, Replicate, and Workers AI behind one API.
HELICONE · TECHHelicone: OSS observability for LLM calls with EU hosting
Helicone is an Apache-2.0 proxy plus cloud (EU region Frankfurt) for LLM cost tracking, caching, and tracing. Setup in under ten minutes.
TRUEFOUNDRY · TECHTrueFoundry: ML platform with embedded LLM gateway
TrueFoundry combines model serving, inference, and LLM gateway in one platform. Self-host (Kubernetes) or cloud, primarily for ML teams with pipelines.
MARTIAN · TECHMartian: model router with embedding classifier per request
Martian (withmartian.com) is a US cloud router that picks the best LLM per request via embedding comparison. Experimental, USD 0.50-2 per 1M tokens surcharge.
BIFROST · TECHBifrost: Go-based self-host LLM gateway under 5 ms overhead
Bifrost (github.com/maximhq/bifrost) is an OSS LLM gateway in Go, self-host, v0.5+ as of May 2026, ultra-low latency for streaming and voice bots.
APISIX AI · TECHApache APISIX AI: OSS API gateway with LLM plugins (ai-proxy, decorator, rate-limiting)
Apache APISIX v3 is an Apache-2.0 API gateway with ai-proxy, ai-prompt-decorator, and ai-rate-limiting plugins. Self-host, Kubernetes, or bare metal.
LANGFUSE · TECHLangfuse: OSS market leader for LLM tracing, prompt management, and eval
Langfuse (MIT, v3+) is the OSS standard tool for LLM tracing, cost tracking, prompt versioning, and eval. Self-host or EU cloud Frankfurt.
MAKE · TECHMake (formerly Integromat): visual scenarios with 1,500+ apps and EU hosting region
Make is a US/EU cloud platform for visual automation scenarios, 1,500+ apps, and operations-based pricing from USD 10.59/10k ops.
ZAPIER · TECHZapier: industry default with 7,000+ apps, US cloud, and task-based pricing
Zapier is the best-known no-code workflow vendor with 7,000+ app integrations, US cloud only, and a Starter tier from USD 19.99/month for 750 tasks.
PIPEDREAM · TECHPipedream: code-first workflows with Python+JS, OSS core, and cloud sourcing
Pipedream pairs cloud convenience with code-first workflows in Python and JavaScript; Free tier 100k credits, Basic from USD 19/month, core on GitHub.
ACTIVEPIECES · TECHActivepieces: MIT-licensed open-source alternative to Zapier, self-hostable for SMEs
Activepieces is an MIT-licensed workflow platform with 250+ pieces, full self-host option, and cloud tier from USD 8/month.
WINDMILL · TECHWindmill: script-first workflows with TypeScript, Python, Go, and Bash under AGPLv3
Windmill is an AGPLv3 platform for code-first workflows in TypeScript, Python, Go, and Bash with approval flows, scheduling, and cloud/self-host.
TEMPORAL · TECHTemporal: durable execution for mission-critical workflows in Java, Go, TS, Python
Temporal is an MIT-licensed workflow-as-code platform with guaranteed retry, state, and versioning – for order processing, payment reconciliation, and more.
APACHE AIRFLOW · TECHApache Airflow: Python DAGs as the data-engineering default since 2014, May 2026 v3.x
Apache Airflow is the standard platform for data pipelines with Python DAGs, Apache 2.0, self-hostable or via Astronomer/MWAA as managed service.
NODE-RED · TECHNode-RED: visual flow editor for IoT, IIoT, and Industry 4.0 with native MQTT support
Node-RED is an Apache 2.0 platform for flow-based programming, originally from IBM 2013, IoT/IIoT default with MQTT, Modbus, OPC-UA.
HUGINN · TECHHuginn: older Ruby-on-Rails tool for personal automation, RSS, scraping, notifications
Huginn is an MIT-licensed Ruby-on-Rails platform for personal automation, IFTTT alternative since 2013, niche but stable.
INNGEST · TECHInngest: event-driven workflows for TypeScript and Python with durable execution
Inngest is an Apache 2.0 SDK plus cloud platform for event-driven workflows with step functions, retry, sleep, and wait-for-event in TypeScript/Python.
INFOMANIAK · TECHInfomaniak: Swiss hosting in Geneva, ISO 27001 + 14001, 100% hydro power, CHF pricing
Infomaniak runs its own data centres in Geneva, keeps 100% Swiss hosting under Swiss law, holds ISO 27001 + ISO 14001 certification and invoices in CHF.
EXOSCALE · TECHExoscale: Swiss cloud from Zurich and Lausanne, ISO 27001, A100 GPU available, FINMA-suitable
Exoscale operates cloud regions in Zurich, Lausanne, Vienna, Frankfurt and Sofia. ISO 27001, DPA available, GPU workloads with A100 in Swiss regions.
OVHCLOUD · TECHOVHcloud: French hoster with four ISO certifications, SecNumCloud, A100/H100 options
OVHcloud operates data centres in Roubaix, Strasbourg, Frankfurt and London. ISO 27001/27701/27017/27018, SecNumCloud-certified, GPU with A100/H100.
SCALEWAY · TECHScaleway: French hoster from Paris, ISO 27001, strong API, NVIDIA H100 AI cluster
Scaleway operates data centres in Paris, Amsterdam and Warsaw. ISO 27001, S3-compatible object storage, H100 cluster for AI training with transparent hourly pricing.
SWISSCOM CLOUD · TECHSwisscom Cloud: Swiss telco hoster, ISO 27001 + FINMA-certified, for banks and insurers
Swisscom Cloud is the cloud arm of Swisscom AG. Data centres in Bern and Zurich, ISO 27001, FINMA outsourcing certification, multilingual support.
SAFE SWISS CLOUD · TECHSafe Swiss Cloud: specialised Swiss hoster in Embrach, FINMA-suitable, fiduciary focus
Safe Swiss Cloud operates a Swiss data centre in Embrach (ZH). ISO 27001, FINMA-compliant, German-language customer service, focus on Swiss SMEs and fiduciaries.
RUNPOD · TECHRunPod: GPU cloud with EU-Sweden region, A100 from USD 1.69/h, H100 from USD 2.59/h
RunPod operates GPU cloud in Hong Kong, USA and EU-Sweden. A100-80GB from USD 1.69/h, H100 from USD 2.59/h. Secure Cloud and Community Cloud tiers.
VAST.AI · TECHVast.ai: GPU auction marketplace, RTX 4090 from USD 0.31/h, A100 from USD 0.40/h, globally distributed
Vast.ai is an auction marketplace for GPU hours, with private and professional providers worldwide. Cheapest spot GPUs, without SLA and with fluctuating availability.
LAMBDA LABS · TECHLambda Labs: dedicated GPU cloud in the USA, A100-40GB from USD 1.10/h, H100 from USD 2.49/h
Lambda Labs is a US GPU cloud with ML engineer focus. A100-40GB from USD 1.10/h, H100 from USD 2.49/h, reserved contracts with 1-year/3-year rebates.
COREWEAVE · TECHCoreWeave: enterprise GPU cloud USA + UK + Spain, H100 from USD 4.25/h, NVIDIA premium partner
CoreWeave is a US GPU cloud focused on enterprise workloads. Sites in US East/West, UK and Spain. H100 from USD 4.25/h, H200 from USD 4.50/h, large clusters with InfiniBand.
BGE-M3 · TECHBGE-M3: open-source embeddings for multilingual RAG systems
BGE-M3 from BAAI is the strongest freely available embedding model in May 2026 for Swiss SMEs. Apache 2.0, 1024-dim, 100+ languages.
MULTILINGUAL-E5 · TECHmultilingual-e5: fast open-source embedding model for CPU setups
Microsoft multilingual-e5 is an mDeBERTa-based embedding model under MIT licence, very fast on CPU and available in four sizes.
VOYAGE AI · TECHVoyage AI: specialised embedding API for RAG setups
Voyage AI is an embedding provider acquired by MongoDB in 2024. voyage-3 costs USD 0.06 per 1M tokens and ranks among the strongest RAG models in May 2026.
JINA EMBEDDINGS · TECHJina Embeddings v3: Berlin-based embeddings with EU cloud and self-host
Jina Embeddings v3 is a multilingual Apache 2.0 model with 8192 token context, operated from Berlin and Frankfurt – EU data protection by default.
MISTRAL EMBED · TECHMistral Embed: EU-native embedding model from Paris
Mistral Embed comes from Paris, costs EUR 0.10 per 1M tokens, and is in May 2026 the politically cleanest API embedding for EU-AI-Act-strict mandates.
NOMIC EMBED · TECHNomic Embed: locally runnable open-source embedding model
Nomic Embed v2 is an Apache 2.0 model with 768 dimensions, transparent training-data documentation, and excellent local performance on Mac and Linux.
MXBAI-EMBED · TECHmxbai-embed: compact Apache 2.0 embedding model for edge setups
mxbai-embed-large-v1 from Mixedbread AI is a 1024-dim open-source embedding, ONNX-capable and fast on edge hardware.
COHERE RERANK · TECHCohere Rerank: industry standard for RAG re-ranking
Cohere rerank-multilingual-v3.0 is the established API reranker for RAG pipelines in May 2026. USD 2 per 1k queries, EU hosting via AWS Bedrock Frankfurt.
BGE-RERANKER · TECHBGE-Reranker-v2-m3: open-source reranker for multilingual RAG setups
BGE-Reranker-v2-m3 from BAAI is the strongest freely available cross-encoder reranker in May 2026 – Apache 2.0, multilingual, very close to Cohere quality.
FLASHRANK · TECHFlashRank: ultra-fast reranker on CPU via ONNX runtime
FlashRank is an MIT-licensed Python library with small cross-encoder models, ONNX-optimised, under 100 ms on CPU.
VLLM · TECHvLLM: production serving for open-weight LLMs with high throughput and PagedAttention
vLLM is an Apache 2.0 inference server for Linux with GPU. PagedAttention and continuous batching deliver up to 20x more throughput than hobby runtimes.
LLAMA.CPP · TECHllama.cpp: the portable C/C++ inference library under Ollama, LM Studio and KoboldCpp
llama.cpp is the MIT-licensed base library for local language models. Runs on every platform – CPU, CUDA, Metal, ROCm, Vulkan. GGUF format standard.
LM STUDIO · TECHLM Studio: desktop app for local LLMs on Mac, Windows and Linux
LM Studio is a graphical desktop app for exploring local open-weight models. Hobby-grade, good for demos and pilot phases, not for multi-user production.
LOCALAI · TECHLocalAI: OpenAI-API-compatible all-rounder for LLM, TTS, STT and vision in one box
LocalAI is an MIT-licensed self-hosting server that bundles LLM, image, audio and embeddings under one OpenAI API. Bare-metal or Docker.
TGI · TECHText Generation Inference (TGI): production serving from the Hugging Face universe
TGI is Hugging Face's Apache 2.0 inference server for production workloads with continuous batching, FlashAttention and direct Hugging Face Hub integration.
LLAMA 4 · TECHLlama 4 Scout and Maverick: Meta's MoE family with 10M context and 400B total parameters
Llama 4 Scout (17B active, 10M context) and Maverick (17B active, 128 experts, 400B total) – released 5 April 2026 under the Llama Community License.
MISTRAL LARGE · TECHMistral Large 2 and Mistral Small 3.1: the EU model pair with FR/DE/IT strength
Mistral Large 2 (123B) and Mistral Small 3.1 from French Mistral AI. La Plateforme, Azure EU, AWS Bedrock. Research or commercial licence.
QWEN 3 · TECHQwen 2.5 and Qwen 3: Alibaba's open-weight family with maths and code strength
Qwen 2.5 and Qwen 3 from Alibaba Cloud. Apache 2.0 for smaller models. Strongly multilingual, leading in maths and code. Self-host resolves the PRC concern.
DEEPSEEK · TECHDeepSeek (V and R lines): the Chinese MoE reasoning model with self-host option
The DeepSeek V and R models under the DeepSeek License. Self-host via Hugging Face. Very strong reasoning, very cheap to operate. API-use warning for client data. Model versions change continuously – verify current names before use.
APERTUS · TECHApertus as a tool: operating the Swiss LLM from ETH, EPFL and CSCS in practice
Apertus 8B and 70B under Apache 2.0. Self-host, Swisscom API or Hugging Face. 15T tokens, 1000+ languages including Swiss German and Romansh.
PROMETHEUS · TECHPrometheus: CNCF time-series DB for metrics, pull model, and PromQL
Prometheus 3.x as CNCF-graduated industry standard for metrics. Pull model, PromQL, service discovery. Self-host, Apache 2.0, SME-ready.
LOKI · TECHLoki: log aggregation as the SME-friendly alternative to Elasticsearch
Loki 3.x from Grafana Labs as the companion log system. Indexes only labels, not full text. AGPL-3, self-host or Grafana Cloud. May 2026 with bloom filters.
UPTIME KUMA · TECHUptime Kuma: SME uptime page with HTTP, TCP, ping, and Docker checks
Uptime Kuma as a self-hosted uptime page. MIT licence, setup in 5 minutes, 13+ monitor types, public status page, SME favourite May 2026.
NETDATA · TECHNetdata: lightweight per-host agent with polished default dashboards
Netdata as per-host agent with hybrid cloud aggregation. GPL-3 plus cloud tier, footprint under 100 MB, polished out-of-the-box dashboards. SME favourite.
SIGNOZ · TECHSigNoz: OpenTelemetry-native APM with metrics, logs, and traces in one
SigNoz as mature open-source alternative to Datadog. MIT licence, OpenTelemetry-first, ClickHouse backend, self-host or cloud. May 2026 SME pick.
GITEA · TECHGitea: lightweight self-host Git server with Gitea Actions
Gitea 1.22 as the SME default for self-hosted Git. MIT licence, single binary, Gitea Actions integrated, container registry, mature and stable May 2026.
GITLAB · TECHGitLab: complete DevOps platform with CI/CD, registry, and security
GitLab 17 as the all-in-one DevOps platform. Community Edition MIT, Enterprise proprietary. Self-host or cloud. Very comprehensive but resource-hungry.
DRONE CI · TECHDrone CI: container-native CI with reduced community activity
Drone CI as a container-native CI system. Apache 2.0, self-host. May 2026: stabilised, reduced development activity. Migration path to Woodpecker.
WOODPECKER CI · TECHWoodpecker CI: active Drone fork as the SME standard for Gitea and Forgejo
Woodpecker CI 3.x as a very active Drone fork. Apache 2.0, container-native YAML pipelines, May 2026 SME default CI for self-host alongside Gitea/Forgejo.
JENKINS · TECHJenkins: the CI veteran with 2000+ plugins, high learning curve, low modernity
Jenkins as the oldest CI server (since 2011). MIT licence, Java-based, 2000+ plugins. Very powerful but hard to maintain. May 2026 not recommended for greenfield.
WHISPER · TECHWhisper: OpenAI open-source STT model for multilingual transcription
Whisper is OpenAI MIT-licensed speech-recognition model. Runs locally via whisper.cpp, faster-whisper or WhisperX, or via API at USD 0.006/min. As of May 2026 with large-v3 and turbo-v3.
DEEPGRAM · TECHDeepgram: proprietary STT API with the lowest latency in the market
Deepgram offers speech-to-text as a US cloud API at USD 0.0043/min Nova-2 with sub-300 ms latency. Strong on English, weak on Swiss German, no EU tier in May 2026.
ELEVENLABS · TECHElevenLabs: the industry reference for natural TTS voices and voice cloning
ElevenLabs offers a proprietary TTS cloud API with the most natural voices in May 2026. Starter USD 5/Mo, Creator USD 99/Mo. turbo-v2.5 for sub-400 ms telephony latency, 30+ languages, voice cloning available.
PIPER TTS · TECHPiper: the open-source local TTS system for privacy-sensitive applications
Piper is an MIT-licensed local text-to-speech system on ONNX. Free, fully local, very fast on CPU. May 2026 with good German voices, hobby-grade quality, good for internal tools.
TWILIO · TECHTwilio: the global telephony standard for programmable voice and SMS applications
Twilio offers programmable telephony and SMS as an API. US headquarters with EU region (Ireland, Frankfurt) for data residency. CH landline minute USD 0.0085. Standard for voice agents.
LANGCHAIN · TECHLangChain: the industry default framework for LLM applications, with all strengths and weaknesses
LangChain in May 2026 in v0.4+ is the most-used LLM framework. MIT license, Python and JavaScript, hundreds of integrations. Strong at prototyping, criticised for code quality and stability.
LLAMAINDEX · TECHLlamaIndex: the clean RAG framework for code-first teams
LlamaIndex in May 2026 in v0.10+ is the cleanest RAG framework. MIT license, Python and TypeScript, clearer API than LangChain. With LlamaCloud also available as managed tier in May 2026.
HAYSTACK · TECHHaystack: the enterprise RAG framework from deepset in Berlin
Haystack in May 2026 in v2.x is the enterprise-focused RAG framework. Apache 2.0, Python, deepset AG (DE origin). Clean pipeline abstraction, commercial support, bank-grade.
DSPY · TECHDSPy: programming instead of prompting – the Stanford approach to LLM pipelines
DSPy in May 2026 in v2.5+ is an MIT framework from Stanford. Instead of writing prompts, you define tasks – the system optimises prompts automatically. Production-capable for complex multi-step pipelines.
RAGFLOW · TECHRAGFlow: the self-hostable all-in-one RAG system with web UI
RAGFlow in May 2026 in v0.15+ is an open-source all-in-one RAG system from Infiniflow (Apache 2.0). Self-host, ready web UI, document parsing, chunking, vector DB, chat. Chinese origin, fully self-hostable.
POSTMARK · TECH STACKPostmark: highest deliverability for transactional emails from the USA
Postmark (ActiveCampaign/Wildbit) is the specialist for transactional mail only. EUR 15/month for 10,000 mails, deliverability > 99 percent, inbound routing included. US hosting, TIA required.
MAILGUN · TECH STACKMailgun: API-strong sending with EU region Frankfurt for developer teams
Mailgun (Sinch) is the developer classic for transactional email. EUR 35/month for 50,000 mails, strong API with SDKs in 8 languages, EU region Frankfurt for GDPR setups, webhooks + routing.
AMAZON SES · TECH STACKAmazon SES: USD 0.10 per 1,000 mails, EU regions incl. Zurich
AWS Simple Email Service is the price winner for sending APIs. USD 0.10 per 1,000 mails, EU regions Frankfurt/Ireland/Zurich, no dashboard out of the box, integrated with the AWS stack.
SENDGRID · TECH STACKSendGrid (Twilio): classic market leader for transactional and marketing email
SendGrid has been the US market leader since 2009. USD 19.95/month for 50,000 mails, EU region available, broad Twilio/Salesforce/HubSpot integration, marketing plus transactional streams.
RESEND · TECH STACKResend: developer-first mail API with React Email templates
Resend is the new developer-first mail API from the React Email makers. Free 3,000 mails/month, Pro USD 20 for 50,000 mails, US hosting with EU region in beta since May 2026.
MAILJET · TECH STACKMailjet (Pathwire/Sinch): EU-first sending from Paris for GDPR setups
Mailjet is the EU-first counterpart to Mailgun. Hosting in Paris, EUR 15/month for 15,000 mails, GDPR + DPA in contract, same Sinch group as Mailgun. Ideal for Swiss SMEs without TIA.
INFOMANIAK · TECH STACKInfomaniak Mail Sender: pure Swiss hosting for fiduciary professional secrecy
Infomaniak is the only provider with pure Swiss hosting in Geneva and Sion. CH-DSG and revDSG compliant, combined with hosting and DNS, perfect for Swiss fiduciaries and law firms.
SMTP2GO · TECH STACKSMTP2GO: multi-region sending from New Zealand, EU and USA
SMTP2GO is a pragmatic mail-sending platform with multi-region routing. USD 15/month for 10,000 mails, clear dashboards, simple setup, EU region selectable for GDPR setups.
MAILERSEND · TECH STACKMailerSend: template editor and solid API with EU region
MailerSend (Lithuanian MailerLite spin-off) is a modern mail-sending platform with drag-and-drop template editor. USD 30/month for 50,000 mails, EU region Frankfurt, clean API, solid DX.
POSTGRESQL · TECHPostgreSQL: the relational default database for Swiss SMEs and AI stacks
PostgreSQL 17 is the industry-default database in May 2026: JSON, full-text search, pgvector, and PostGIS in one system. MIT-like license, self-hostable.
MYSQL & MARIADB · TECHMySQL and MariaDB: the classic LAMP stack honestly assessed for 2026
MySQL (GPL-2, Oracle) and MariaDB (BSL/GPL-2, MariaDB Foundation) are the LAMP classics. In May 2026, MariaDB 11 is stable, a serious OSS alternative to MySQL 8.4.
SQLITE · TECHSQLite: the single-file database for single-tenant, mobile, and edge
SQLite is a public-domain embedded DB as a single file. In May 2026 in production at Fairlane and Realty51, with Litestream replication to S3. Very fast, very robust.
MONGODB · TECHMongoDB: the document database between SSPL, Atlas, and Voyage AI integration
MongoDB is the best-known document DB. SSPL license, self-host or Atlas Cloud. May 2026 with Voyage AI integration for native embedding and vector search.
REDIS · TECHRedis as cache layer: KV store, sessions, rate limits, pub/sub
Redis 8 is the KV and cache standard in May 2026. SSPL since 2024 (alternatively Valkey under BSD). Very fast, microsecond latency, never as primary DB.
CLICKHOUSE · TECHClickHouse: columnar analytics database for aggregates over billions of rows
ClickHouse 25 is in May 2026 the fastest OSS columnar DB for OLAP workloads. Apache 2.0, self-host or cloud. 100x faster than Postgres on aggregates.
COCKROACHDB · TECHCockroachDB: distributed Postgres-compatible SQL for multi-region setups
CockroachDB is a distributed ACID DB with Postgres wire protocol. BSL with Apache 2.0 conversion after 3 years. Self-host or cloud, multi-region capable.
SURREALDB · TECHSurrealDB: multi-model database in Rust with document, graph, and time-series
SurrealDB 2.x is stable in May 2026. BSL with 4-year Apache 2.0 conversion, Rust-based, multi-model relational + document + graph + time-series.
DUCKDB · TECHDuckDB: embedded columnar OLAP database for local data analysis
DuckDB 1.x is stable in May 2026. MIT license, embedded, columnar, perfect for local data analysis -- Pandas replacement and fiduciary reporting tool.
SUPABASE · TECHSupabase: Postgres-based backend-as-a-service with EU region Frankfurt
Supabase is Apache 2.0 open-source Postgres with auth, storage, realtime, and edge functions. EU cloud Frankfurt or self-host. Fast deployment for MVPs.
CROWDSEC · TECHCrowdSec: open-source WAF with collaborative blocklist for SME servers
CrowdSec is the MIT-licensed intrusion detection with crowd-sourced threat intelligence from 100k+ servers. May 2026 SME market leader with AI Bouncer.
FAIL2BAN · TECHFail2ban: classic log-based IP banning for Linux servers
Fail2ban is the GPL-2 classic for intrusion detection since 2004. Simple, stable, no crowdsource layer. May 2026 v1.x stable, predecessor of CrowdSec.
WAZUH · TECHWazuh: SIEM, EDR and compliance platform for the regulated mid-market
Wazuh is the GPL-2 SIEM platform with file integrity, vulnerability detection, and PCI-DSS/HIPAA reports. May 2026 the pick for 50+ staff law firms.
HASHICORP VAULT · TECHHashiCorp Vault: industry standard for secrets management since 2015
Vault is the market leader for secrets management. Since 2023 under BSL 1.1 (no longer MPL-2). Self-host and cloud. Very powerful, setup 5-15 days.
OPENBAO · TECHOpenBao: Linux Foundation fork of HashiCorp Vault under MPL-2.0
OpenBao is the MPL-2.0 fork of Vault under Linux Foundation governance. May 2026 v2+ active, direct replacement for Vault BSL migration.
AUTHELIA · TECHAuthelia: lightweight SSO and 2FA authorization proxy for nginx and Traefik
Authelia is the Apache 2.0 open-source solution for SSO+2FA+authorization proxy. Go-based, below 30 MB footprint, very SME-friendly.
AUTHENTIK · TECHAuthentik: modern identity provider system with SAML, OIDC, LDAP and SCIM
Authentik is the MIT-licensed IdP system (plus Enterprise) for SAML/OIDC/LDAP/SCIM. Self-host and cloud, May 2026 v2025+ active.
WIREGUARD · TECHWireGuard: modern VPN in the Linux kernel for home office and remote admin
WireGuard is the GPL-2 VPN in the Linux kernel. 4000 lines of code, 3-10x faster than OpenVPN. SME default in 2026 for remote access.
CLAMAV · TECHClamAV: open-source antivirus for mail servers and upload scanning
ClamAV is the GPL-2 antivirus classic for mail server gateways and file upload scanning. Update hub with ClamSubmit community.
BITWARDEN · TECHBitwarden: password manager, secrets manager and passkey authenticator with EU cloud
Bitwarden is the GPL-3 OSS solution for passwords and secrets management. Self-host and EU cloud Frankfurt. Plus Passkey native auth.
Tools · Microsoft 365 CopilotMicrosoft 365 Copilot for Swiss Trustees: Data Residency, Flex Routing & revFADP Checklist
Data residency, EU Data Boundary, Flex Routing, no training on tenant data and revFADP processing rules – what really applies when using Copilot with client data.
Tools · AI Bookkeeping SwitzerlandAccounto, Bexio (Kontera AI) & KLARA: AI Bookkeeping Compared for Fiduciary Firms
Neutral comparison of Swiss AI bookkeeping (Accounto, bexio/Kontera, KLARA, Abacus): receipt recognition, automation, fiduciary collaboration, CH data residency.
Implementation & Rollout
From pilot to production, costs.
AI-Readiness Audit: where your business stands with AI today – clarified in one to five days
Stack scan, data-protection check, three prioritised use cases. Report plus 90-min debrief. Light from CHF 1,700, Deep CHF 3,900.
SERVER & INFRASTRUCTURE · SERVICEServer & Infrastructure: Ubuntu, Docker, monitoring – set up, hardened, handed over
Your own server on Hetzner or DigitalOcean. Ubuntu 24.04, Docker, nginx, SSL, Fail2ban, CrowdSec, monitoring. Fixed price from CHF 1,200.
MULTI-LLM GATEWAY · SERVICEMulti-LLM Gateway: eight providers, one entry point, compliance routing
LiteLLM gateway with auth, routing by cost/speed/privacy, usage dashboard. Module from CHF 1,000, project with observability CHF 4,500.
RAG ON YOUR OWN KNOWLEDGE · SERVICERAG on your own knowledge: answers from your documents – with sources, not made up
Searchable knowledge base with chat. PII redaction, citation check, DE/FR/IT/EN. Pilot up to 10,000 docs CHF 3,500, Production from CHF 8,500.
n8n · SERVICEn8n Workflow Automation: routine out, minds free
Self-hosted n8n on your server plus ready-made workflows. Invoice triage, lead routing, mail classification, daily summary – documented and handed over.
MANAGED · SERVICEManaged Service & Monitoring: we keep it running, you use it
Monitoring, updates, security patches, incident response. Three tiers: Basic CHF 600/mo, Pro CHF 1,200/mo, Plus CHF 2,200/mo. Defined response times.
OWN LLM · COSTSWhat does your own LLM cost? Total cost of ownership in May 2026
Hardware, power, DevOps, maintenance: every TCO building block for a self-hosted language model, with real CHF and USD figures for May 2026.
RAG PILOT · COSTSWhat does a RAG pilot cost? Three tiers: 1k, 10k, 100k documents
A real-world cost breakdown for a RAG pilot in Switzerland: embedding, vector DB, LLM inference, setup effort. Three volume tiers with figures in CHF.
GPU PRICES · COSTSGPU cost calculator 2026: T4, L4, A10, A40, A100, H100, H200 compared
Which GPU fits which model, what does it cost at which provider, on-demand vs. reserved? May 2026 prices from AWS, GCP, Azure, Hetzner, RunPod, Vast.ai.
TOKEN PRICING · COSTSToken costs explained: input, output, cache, provider comparison May 2026
What is a token, how does input pricing differ from output, what does 1M tokens cost at which provider? Table with every relevant model.
BREAK-EVEN · COSTSCloud API vs. self-host: at what token volume does which pay off?
Break-even analysis with May 2026 numbers. 1M / 10M / 100M / 1B tokens per month: where do cloud and self-host curves cross? Plus hidden costs.
COSTS · AI AUTOMATION SMEWhat does AI automation cost for a Swiss SME? Honest prices June 2026
From CHF 590 speedcheck to CHF 20,000 first-year for a 5-FTE fiduciary: all line items, server rent, token costs, managed service.
ROI · AI PROJECTROI calculator for AI projects: how to calculate honestly in May 2026
Formula, Swiss fiduciary collections example (break-even month 3), Frauenfeld case at 7-to-8-week ROI – and the "time saved is not linearly money" pitfall.
STRUCTURED EXTRACTION · AI CONCEPTStructured extraction with grounding: cite-the-source from PDFs and emails
Extract data from PDFs and emails with source proofs: Gemini Citation API, Anthropic with_citations, manual citation linkers and audit-trail link.
BEXIO API · INTEGRATIONBexio API: AI integration into Swiss fiduciary bookkeeping
The Bexio REST API connects Swiss SME bookkeeping with AI workflows. OAuth 2.0, Swiss hosting, a clear data model.
ABACUS API · INTEGRATIONAbacus API: AI connection to the Swiss ERP platform
Abacus web services connect ERP, payroll, and bookkeeping with AI workflows. REST endpoints, SOAP legacy, on-prem or cloud.
SAP B1 · INTEGRATIONSAP Business One: AI integration via DI-API and Service Layer
SAP Business One is SAP's SME ERP. Two APIs: the DI-API (Java/.NET, low-level) and the Service Layer (REST, modern). AI connects via the Service Layer.
SAGE 50 · INTEGRATIONSage 50: AI integration via the Sage Connect REST API
Sage 50 is Sage's SME bookkeeping software in the DACH region. The Sage Connect API was modernised in May 2026 and is now a REST interface with OAuth 2.0.
MS GRAPH · INTEGRATIONMicrosoft 365 Graph API: mail, calendar, Teams, and SharePoint as AI source
The Microsoft Graph API is the central gateway to Microsoft 365. OAuth 2.0, granular permissions, throttling limits. The tool for AI in the Microsoft stack.
GOOGLE WORKSPACE · INTEGRATIONGoogle Workspace: Gmail, Calendar, Drive, and Docs as AI source
Google Workspace APIs provide access to Gmail, Calendar, Drive, Docs, and Sheets. May 2026 GA: AppSheet and Vertex AI Connect for AI workflows.
SHAREPOINT · INTEGRATIONSharePoint and OneDrive: document RAG source for fiduciary and law firm
SharePoint and OneDrive form the document store of most Swiss firms. REST API and Graph for AI indexing, CSOM as legacy.
SSO · INTEGRATIONSSO with SAML 2.0 and OIDC: one login for Bexio, Microsoft 365, and AI apps
SAML 2.0 for enterprise, OIDC for modern apps. IdPs: Entra, Google, Keycloak, Authelia. May 2026: passkeys and SCIM for user provisioning.
WEBHOOKS · INTEGRATIONWebhooks and event-based integration: HMAC, idempotency, retry
Webhooks replace polling with push events. HMAC signature, replay protection, idempotency keys. May 2026 best practices for robust pipelines.
REST · GRAPHQL · INTEGRATIONREST vs GraphQL: which API architecture for AI integrations?
REST dominates May 2026, GraphQL for complex data models. OpenAPI 3.1, MCP server for LLM access. Tools: FastAPI, Express, tRPC.
CHUNKING · AI CONCEPTChunking strategies for RAG: fixed-size, recursive, semantic, late chunking
How to slice documents for RAG: fixed-size, recursive, semantic, document-based and late chunking compared, with rules of thumb for contracts, tables and multilingual texts.
DOCUMENT LOADERS · AI CONCEPTDocument loaders: cleanly ingesting PDF, DOCX, XLSX, EML, HTML and Markdown into RAG
Which tools convert which document formats losslessly into a RAG pipeline: Unstructured.io, LlamaParse, MarkItDown (Microsoft, May 2026), PyMuPDF and pandoc compared head-to-head.
OCR · AI CONCEPTOCR for receipts and contracts: Tesseract, AWS Textract, Azure DI, Google DocAI, Mistral OCR, Reducto
Which OCR engine to choose in May 2026 for scanned contracts, receipts and forms: accuracy, price, EU hosting, Swiss data-protection readiness and use-case recommendations.
PDF TABLES · AI CONCEPTPDF table extraction: Camelot, Tabula, pdfplumber, Table Transformer, Marker
Which tool extracts balance sheets, VAT tables and payslips losslessly from PDFs in May 2026: Camelot, Tabula, pdfplumber, Microsoft Table Transformer and Marker compared head-to-head.
EMAIL ARCHIVES · AI CONCEPTIndexing email archives: IMAP, EWS, Microsoft Graph, MBOX and Swiss data-protection compliance
How to index 5 to 10 years of client correspondence for RAG: IMAP, EWS, Microsoft Graph, MBOX imports, attachment extraction, thread preservation, Swiss-DSG and professional-secrecy compliant.
DATA CLEANING · AI CONCEPTData cleaning before RAG: duplicates, boilerplate, OCR artefacts, charset issues, watermarks
Why 30 percent of any RAG corpus is junk and how to remove it: duplicate detection, header/footer stripping, OCR correction, encoding repair and watermark removal with cleanlab and dedupe.io.
OLLAMA · HOW-TOInstall Ollama: step-by-step guide for Mac, Linux and Windows (May 2026)
Practical guide to installing Ollama 0.5+ on macOS, Linux and Windows including model download, REST API test, Q4_K_M quantisation, systemd setup and GPU acceleration.
RAG · HOW-TORAG pilot in 7 days: from 50 PDFs to a working knowledge base (May 2026)
Day-by-day guide from 50 PDFs to working retrieval-augmented generation with pgvector, BGE-M3, LiteLLM, Streamlit UI and Ragas eval. Budget CHF 800-1500.
N8N · HOW-TOn8n self-host with Bexio integration: from Docker-Compose to a dunning workflow (May 2026)
Guide for self-hosted n8n on EU hosting with Postgres backend, Bexio OAuth app, webhook receiver and a complete workflow "on new invoice check dunning status". Production-ready.
BREVO · HOW-TOSet up Brevo SMTP: DNS, DKIM, DMARC, Nodemailer test and bounce webhooks (May 2026)
Step-by-step guide from DNS records via Hetzner firewall whitelisting to a Nodemailer test, with production tips for dedicated IP and bounce webhook in n8n.
CLOUDFLARE · HOW-TOSet up Cloudflare with your own domain: DNS, SSL, WAF, Workers KV and Tunnel (May 2026)
Guide from domain registration via nameserver change, SSL modes, DNS records, page rules, Workers KV, WAF rules and Cloudflare Tunnel for origin protection.
BACKUP · HOW-TOBackup strategy 3-2-1 for SMEs: restic, rclone, Backblaze B2 and recovery drills (May 2026)
A real-world setup for SMEs with a Linux server, Bexio cloud and Office 365: restic for files, mysqldump for the DB, rclone for cloud, object lock on Backblaze B2. Budget CHF 30-50/month.
LITELLM · HOW-TOInstall the LiteLLM gateway: Docker, config.yaml, virtual keys, cost tracking and Langfuse (May 2026)
Guide from empty server to a production LiteLLM proxy with 5 providers (OpenAI, Anthropic, Mistral, Gemini, Ollama), virtual keys with budget limits, PostgreSQL audit log and Langfuse tracing.
ImplementationPrompt Library for Trustees: Proven Prompts for Accounting, VAT and Correspondence
Safe, reusable prompt patterns for everyday trustee work – with clear limits and no client data in unsafe tools.
Implementation · Client Portal & RAGClient Portal with an AI Assistant: secure self-service chat on your own trustee knowledge (RAG)
Client portal with a RAG chatbot on your own firm knowledge: access control, tenant isolation, revDSG, hallucination limits and escalation to a human.
Risks & Limitations
What goes wrong, what AI cannot do.
Limiting hallucinations: five countermeasures against fabricated AI answers
Why language models produce plausibly wrong answers, which five remedies reduce them, and how to make hallucinations measurable.
SHADOW AI · COMPLIANCEShadow AI in the enterprise: when employees use ChatGPT privately on client data
Cyberhaven 2026: 32 % of ChatGPT use runs through personal accounts. 39.7 % of inputs contain sensitive data. Consequences + countermeasures.
BACKUP · SECURITYBackup strategies 3-2-1 and 3-2-1-1-0: how to secure an SME audit-ready
Three copies, two media, one offsite, one offline, zero verification errors. Tools, cloud targets and Swiss data law aspects as of May 2026.
DISASTER RECOVERY · SECURITYDisaster recovery, RTO and RPO: what an SME fiduciary really must keep ready
How long may the outage last, how much data may be lost? Four DR strategies with May-2026 pricing and tooling.
RBAC · SECURITYRBAC and access management: who may see what in a fiduciary system?
Role-Based Access Control vs ABAC, least privilege, client separation. Tools like Authelia, Authentik, Keycloak and Entra in the SME comparison.
SECRETS · SECURITYSecrets management with Vault: handling API keys, DB passwords and JWT secrets correctly
No .env files in Git, no cleartext backups. HashiCorp Vault, Bitwarden Secrets, 1Password, AWS and Entra in the SME comparison.
PENTEST · SECURITYPentest and vulnerability scans: what an SME must check yearly
Manual pentest, automated scans, bug bounty and OWASP LLM Top 10. Swiss providers, OSS tools and realistic May 2026 prices.
INCIDENT RESPONSE · SECURITY & OPSIncident response playbook: NIST SP 800-61 six-phase model for SMEs
Structured response to security incidents in six phases, with FADP-compliant 72-hour notification to the FDPIC and tooling such as TheHive, Wazuh, and MISP.
LOGGING · SECURITY & OPSLogging and audit trail: audit-proof logging under Swiss CO Art. 957a for SMEs
Immutable logs with 10-year retention under Swiss CO 957a, WORM storage, and hash chains. Tools as of May 2026: Loki, Wazuh, GrayLog, Elasticsearch.
DDoS · SECURITY & OPSDDoS protection with Cloudflare: layer 3, 4, and 7 for SME web apps in 2026
Defend against volumetric and application-layer DDoS attacks with Cloudflare Free, Pro, or Business. As of May 2026 including AI-orchestrated attacks.
TLS · SECURITY & OPSSSL/TLS certificates with Lets Encrypt: TLS 1.3, HTTP/3, and auto-renew for SMEs 2026
Free TLS certificates with 90-day validity, auto-renew via Certbot or Caddy, TLS 1.3 plus HTTP/3 as the 2026 standard.
FIREWALL · SECURITY & OPSFirewall and CrowdSec: layered protection for SME servers in 2026
Network firewall (ufw/nftables) plus application protection with CrowdSec, the open-source successor to Fail2ban with crowd-sourced threat intel.
HALLUCINATION MEASUREMENT · AI CONCEPTDetecting and measuring hallucinations: metrics, benchmarks and self-consistency
How to measure hallucinations in AI answers reproducibly: TruthfulQA, HaluEval, FActScore, self-consistency and citation grounding checks.
EVAL FRAMEWORKS · AI CONCEPTEval frameworks for LLMs: DeepEval, OpenAI Evals, Promptfoo, Ragas, TruLens compared
Which LLM evaluation framework when: DeepEval, OpenAI Evals, Promptfoo, Inspect (UK AISI), Ragas, TruLens, MLflow LLM Evaluate and Phoenix Evals.
GOLDEN DATASET · AI CONCEPTBuilding a golden dataset: 50-500 test examples done right for SMEs
Stratified sampling, edge cases, adversarial set, quarterly refresh and annotation guidelines for a solid test set in everyday fiduciary practice.
LLM-AS-A-JUDGE · AI CONCEPTLLM-as-a-judge: AI evaluates AI – methods, bias pitfalls, limits
GPT-4 and Claude as eval judges, pairwise vs pointwise scoring, position bias and self-preference, the G-Eval paper, when humans remain indispensable.
REGRESSION TESTING · AI CONCEPTRegression testing for LLMs: CI/CD, snapshot tests and detecting model-update drift
CI/CD integration of LLM tests, snapshot testing for prompts, diff testing between model versions using the the current Claude model-to-4.7 fiduciary pipeline example.
RED-TEAMING · AI CONCEPTRed-teaming for AI: jailbreaks, prompt injection and OWASP LLM Top 10 v2.0
Adversarial prompts, DAN mode, prompt injection (direct and indirect), OWASP LLM Top 10 v2.0 and May 2026 tools: PyRIT, Garak, Promptfoo Red-Team.
BIAS & FAIRNESS · AI CONCEPTBias and fairness audits for AI: Swiss equality law, EU AI Act Art. 10, BBQ and StereoSet
How SMEs measure bias in LLM outputs: Swiss equality law, EU AI Act Art. 10, BBQ benchmark, StereoSet, CrowS-Pairs and a fiduciary example on language accent.
AI KPIS · AI CONCEPTMeasuring AI quality: KPIs for RAG, latency, cost and user satisfaction
SME dashboard for AI quality: faithfulness, answer relevancy, context precision, context recall, latency, cost-per-query, user satisfaction.
People & Organisation
Jobs, acceptance, change.
AI for Swiss HR and staffing: CV screening, voicebot interviews and employee FAQ
How HR departments and staffing firms use AI in CV screening, job ads, interview voicebots and employee FAQ – with bias audits, revised FADP and EU AI Act high-risk duties.
People & OrganisationShadow AI in the fiduciary firm: policy, tool approval list & staff training
How fiduciary firms control unsanctioned AI use: internal policy, approval list, training and Swiss data-protection context.
Trends & Context
What is new, what is hype.
Agentic AI trend 2026: what MCP, Computer Use and multi-agent frameworks really deliver
May 2026: 1500+ MCP servers, Computer Use in production, AutoGen 0.4 and CrewAI tested in SMEs. What works and where token costs spiral.
MULTIMODAL · TREND 2026Multimodal LLM trend 2026: image, audio and video as standard inputs
May 2026: GPT-4o, the current top Claude model and Gemini 2.5 Pro read images, hear speech and understand video. What that means in practice for fiduciary and receipt workflows.
EDGE AI · TREND 2026Edge AI trend 2026: on-device models for phone, laptop and client app
May 2026: Apple Intelligence, Phi-4 and Llama 3.2 run locally on devices. What that means for privacy, latency and offline capability in SME apps.
REASONING · TREND 2026Reasoning model trend 2026: o3, R1, Extended Thinking and the test-time-compute boom
May 2026: OpenAI o3, Claude with Extended Thinking, Gemini 2.5 Pro Thinking and DeepSeek-R1. When the several-times token premium is worth it for SMEs.
OPEN-WEIGHT · TREND 2026Open-weight vs closed trend 2026: how close are Llama 4, DeepSeek and Mistral to GPT and Claude?
May 2026: open-weight is closing the gap. The current DeepSeek-V generation matches GPT-4o, Llama 4 Maverick reaches Claude Sonnet. What licence and hosting actually mean.
EU AI ACT · TREND 2026EU AI regulation trend 2026: deadlines, GPAI obligations and Switzerland's response
May 2026: prohibitions in force since February 2025, GPAI models regulated since August 2025, high-risk from August 2026. Swiss link via market access.
ENERGY & CO2 · TREND 2026AI energy and CO2 trend 2026: what a query actually consumes and where Switzerland stands
May 2026: 0.3 Wh per GPT-4 query, 500t CO2 for GPT-4 training, data centres up 35% YoY. Swiss advantage: hydropower and waste-heat use.
Trends & ComparisonTREUHAND|SUISSE GPT vs. Your Own AI Infrastructure: Data Protection, Vendor Lock-in, Cost
Neutral comparison: the association tool TREUHAND|SUISSE GPT versus a vendor-neutral multi-LLM gateway with your own RAG on firm data.