Data, Knowledge, and Retrieval for Agents
What this domain tests
Data, Knowledge, and Retrieval for Agents tests whether you can build grounded answers: ingestion, chunking, embeddings, vector search, metadata filters, freshness, and citation. It also tests access control at retrieval time.
Core rules
- RAG has two paths: offline ingestion (parse, chunk, embed, index) and online retrieval (embed query, search, rerank, assemble context). Failures on either path look like bad answers.
- Chunk by meaningful structure, such as headings or policy sections, not fixed character counts alone. Preserve parent-child links so a retrieved chunk can be expanded.
- Store metadata with every chunk: source, version, effective date, region, classification, and allowed roles. Use it to filter before similarity search, not after.
- Retrieval quality is not just top-k. Use hybrid search when exact terms matter, rerank candidates, and set a similarity threshold so the agent can say not found.
- Grounding requires citations and a refusal path. The agent should not answer from parametric memory when the knowledge base is the authority.
Worked example
A support agent answers questions about product warranties. Documents are chunked by warranty section, and each chunk carries product line, region, and effective date. A query for EU battery warranty retrieves only chunks where region=EU and effective_date is current. The agent must cite the section and, if no chunk clears the threshold, escalate. The reasoning is that metadata filtering prevents a correct-sounding answer from the wrong region, and citation makes the answer auditable.
Traps
Do not embed documents without access metadata and then filter in the prompt. Do not assume a larger top-k improves answers; it adds noise. Do not ignore document versioning, or the agent will cite superseded policy. Do not let the model invent a citation.