RAG

RAG is the pipeline that turns documents into grounded answers: parse and chunk sources, embed or index them, retrieve relevant passages for a query, then generate with those passages in context (ideally with citations). It lives under Agentic because the product job is an agentic/knowledge loop, not raw model serving. It is not Vector DBs (those are the ANN substrate many RAG stacks call) and not Memory (session/profile persistence), though production systems often compose all three.

Technically the hard parts are corpus hygiene and retrieval quality: OCR and layout-aware parsing, chunk boundaries that preserve meaning, hybrid search, reranking, and citation fidelity. Failures look like fluent answers with wrong sources, silent empty retrieval, or indexes that never update when Monday-style boards get a typed row. In a company OS, RAG is how invisible requests and buried email become countable evidence — if the ingest path sees them.

Members in this cell mix full stacks and document engines:

• LlamaIndex and Haystack are orchestration-shaped RAG frameworks: connectors, indexes, query pipelines.
• RAGFlow and txtai lean toward packaged retrieval+generate products.
• Unstructured, Docling, markitdown, MinerU, and PaddleOCR attack the parse/OCR wall that kills grounding before any vector query runs.

Prefer a full framework when you own the loop end-to-end; prefer parse specialists when the bottleneck is PDFs and scans, not the LLM. Wire retrieval so every cited span maps to a durable source id. Feature maps already include Knowledge Grounding & RAG and Citations — treat those as the falsifiers: can you show the passage that justified the claim?

Navigate: use Vector DBs when the missing piece is embedding storage/ANN; Knowledge Graphs when typed relations beat cosine; Evaluations when you need to score retrieval/generation quality. Stay in RAG when the missing piece is grounding generation in a corpus you control.

Retrieval quality is falsifiable: freeze a query set, record hit@k and citation precision, and refuse to “improve the prompt” when the index is stale. Chunking strategy, embedder choice, and filter metadata are product decisions — they decide whether an estimator’s invisible email becomes a countable request or stays dark. Prefer hybrid retrieval when proper nouns and SKUs matter; prefer rerankers when the first-pass ANN list is noisy. Never treat the LLM’s confidence as a substitute for a retrieved span you can open.

For this research catalog the acceptance test is boring and strict: after wiring the pack, can a second person replay the same scenario and get the same measurable outcome? If not, the page is still a brochure. Document the scenario, the expected number, and the log line that proves it. That habit is what separates a deepened taxonomy cell from a landscape screenshot.