Last reviewed: 2026-10. Focus: EU / GDPR.
Intro & scope
The KNOW family covers AI that finds, answers from, curates and researches information. Four patterns: - KNOW.SEARCH - retrieval first, a ranked hit list (enterprise, product, multimodal, archive, CAD, spare-part, tender search). - KNOW.QA - cited answers over a bounded corpus (company docs, HR, technical, standards, clinical). - KNOW.RESEARCH - multi-step research over external corpora (legal, patent, science, trademark, finance, OSINT, deep research). - KNOW.MANAGE - building and maintaining the knowledge itself (KB/FAQ, expert finding, graph, memory, records, SOPs, personal notes).
Scope boundaries between overlapping files: - KNOW-001 (search, hit list first) vs KNOW-002 (generated answer first): same index, different UI and risk; 002 is the answering layer on 001. - KNOW-002 is the generic company-docs assistant; KNOW-003, 004, 009, 010 are domain-specialised variants (HR, service technicians, standards, clinical) with their own sources and duties. - KNOW-006/007/022 are IP and legal research on external registers; KNOW-015 is open-web desk research; KNOW-008 is scientific literature; KNOW-023 is financial corpora. - KNOW-013 builds the graph (data layer); KNOW-026 uses graphs for investigations (people, high risk). KNOW-025 manages records and retention, KNOW-011 manages article content. - KNOW-004 answers manual questions, KNOW-021 identifies the part, KNOW-020 finds similar CAD parts, KNOW-005 is shop search with a catalogue. Cross-family: document processing in DOC (esp. DOC-014), agents in AGENT, speech in SPEECH, vision in VIS.
Path legend: A = managed/SaaS or API build, B = assembled platform/partner build, C = full self-host (HW tier is the minimum sensible tier for path C).
Use cases
| ID | Title | Pattern | Maturity | Path | HW tier | AI Act risk |
|---|---|---|---|---|---|---|
| KNOW-001 | Enterprise semantic search across intranet, drives & tools | KNOW.SEARCH | mainstream | A | HW2 | minimal |
| KNOW-002 | RAG assistant over company documents ("chat with your data") | KNOW.QA | mainstream | A | HW2 | limited |
| KNOW-003 | HR & policy self-service assistant for employees | KNOW.QA | mainstream | A | HW2 | limited |
| KNOW-004 | Technical documentation & service manual assistant for technicians | KNOW.QA | early-adopter | A | HW3 | minimal |
| KNOW-005 | E-commerce semantic product search | KNOW.SEARCH | mainstream | A | HW1 | minimal |
| KNOW-006 | Legal research assistant (case law, statutes, commentary) | KNOW.RESEARCH | early-adopter | A | HW3 | depends |
| KNOW-007 | Patent search & prior-art analysis | KNOW.RESEARCH | early-adopter | A | HW3 | minimal |
| KNOW-008 | Scientific literature review & research assistant | KNOW.RESEARCH | mainstream | A | HW3 | minimal |
| KNOW-009 | Standards & regulatory compliance Q&A (ISO, DIN, norms) | KNOW.QA | early-adopter | B | HW2 | minimal |
| KNOW-010 | Clinical guideline & medical knowledge assistant | KNOW.QA | early-adopter | A | HW3 | depends |
| KNOW-011 | Automatic knowledge base & FAQ generation and maintenance | KNOW.MANAGE | early-adopter | A | HW2 | minimal |
| KNOW-012 | Expert finding & skills mapping | KNOW.MANAGE | early-adopter | A | HW2 | depends |
| KNOW-013 | Knowledge graph construction from documents (GraphRAG) | KNOW.MANAGE | early-adopter | B | HW2 | minimal |
| KNOW-014 | Organisational memory & lessons-learned capture | KNOW.MANAGE | early-adopter | B | HW2 | depends |
| KNOW-015 | Deep-research agent for market & desk research | KNOW.RESEARCH | mainstream | A | HW3 | minimal |
| KNOW-016 | Multimodal search in image, video & audio archives | KNOW.SEARCH | early-adopter | A | HW2 | depends |
| KNOW-017 | Public tender & funding opportunity discovery and matching | KNOW.SEARCH | mainstream | A | HW1 | minimal |
| KNOW-018 | Personal knowledge management ("second brain") | KNOW.MANAGE | mainstream | A | HW1 | minimal |
| KNOW-019 | Archive, library & cultural-heritage collection discovery | KNOW.SEARCH | early-adopter | B | HW3 | minimal |
| KNOW-020 | 3D CAD similar-part search & part reuse (Ähnlichteilsuche) | KNOW.SEARCH | early-adopter | A | HW2 | minimal |
| KNOW-021 | Spare-part identification & aftersales parts search | KNOW.SEARCH | early-adopter | A | HW2 | minimal |
| KNOW-022 | Trademark & design clearance search and infringement watch | KNOW.RESEARCH | mainstream | A | HW2 | minimal |
| KNOW-023 | Investment & financial research assistant (filings, earnings calls, broker research) | KNOW.RESEARCH | mainstream | A | HW3 | minimal |
| KNOW-025 | Records management: auto-classification, retention labelling & ROT clean-up | KNOW.MANAGE | early-adopter | A | HW2 | minimal |
| KNOW-026 | Investigative intelligence & OSINT analysis (entity resolution, link analysis) | KNOW.RESEARCH | early-adopter | C | HW3 | depends |
| KNOW-027 | SOP & work instruction generation from screen recordings and expert videos | KNOW.MANAGE | mainstream | A | HW2 | minimal |
Shared stack
Principle: one retrieval platform per organisation, reused by all KNOW use cases (and DOC, AGENT, CONV). Differences between use cases are corpora, ACLs, prompts and evaluation sets, not infrastructure.
Common layer (all tiers)
- Ingestion: connectors (SharePoint, Confluence, file shares, ticketing) plus a document parser/OCR (Docling, Unstructured, or Azure/Mistral OCR) producing structured chunks with ACL metadata and source page references.
- Index: PostgreSQL + pgvector (or Qdrant / OpenSearch / Elasticsearch) holding BM25 and vector side by side (hybrid search); ACLs applied as query filters.
- Retrieval: multilingual embedding (e.g. bge-m3, multilingual-e5, Qwen3-Embedding) and a cross-encoder reranker (bge-reranker-v2-m3 or Cohere Rerank via EU region).
- Answering: LLM with mandatory citations, refusal when no source, guardrails (prompt-injection, PII, output checks), logging with retention limits.
- Ops: evaluation set per use case (retrieval recall, faithfulness), identity via SSO, EU region hosting, DPA and no-training clauses with any API vendor.
HW1 (CPU / edge, < 16 GB)
Uses: KNOW-005, 017, 018. - Postgres + pgvector or OpenSearch on CPU; small embedding model (bge-m3 or e5-small via ONNX/TEI on CPU); reranker on CPU for top-50. - Generation via EU-hosted API or a 3-8B model through Ollama/llama.cpp where needed; personal use (KNOW-018) runs fully local (laptop NPU).
HW2 (one 16-32 GB GPU)
Uses: KNOW-001, 002, 003, 009, 011, 012, 013, 014, 016, 020, 021, 022, 025, 027. - Inference node (vLLM or Ollama) with a quantised 14-30B instruct model (Qwen3, Mistral Small, Gemma 3), embeddings and reranker on the same GPU via TEI. - Qdrant or pgvector, plus OpenSearch for BM25. Add a small VLM (Qwen2.5-VL 7B) and Whisper (large-v3-turbo) for KNOW-016, 021, 027; Neo4j/Memgraph for KNOW-013 and 012; classifier fine-tunes for KNOW-025. - Share one node across the HW2 use cases through an LLM gateway (LiteLLM) with per-tenant quotas and logging.
HW3 (one 48-141 GB GPU, or Mac Studio class)
Uses: KNOW-004, 006, 007, 008, 010, 015, 019, 023, 026. - Everything from HW2, plus a 70-120B class model (gpt-oss-120b, Llama 3.3 70B, Qwen3-235B quantised) or reasoning model for long multi-document synthesis; long-context serving through vLLM. - Agent layer (LangGraph or similar) for deep research with web fetch in a sandbox; larger VLM for scans and drawings; graph DB with community summarisation for GraphRAG (KNOW-013, 019, 026). - Sensitive corpora (clinical, investigative, legal) stay on path C; regulated external corpora (patents, financial data) are licensed and kept as separate indexes.
Sharing with other families
- DOC: reuse the parsing/OCR pipeline and chunk store (DOC-014); do not build a second one.
- AGENT / CONV: expose retrieval as a tool (MCP or REST) with the same ACL filter instead of private copies of the index.
- SPEECH / VIS: transcription and image embeddings feed the same index as extra modalities.
- Move up a tier only when a measured quality gap justifies it; most HW1-HW2 use cases can start on an EU-hosted API (path A) with the same index and swap the generator later.