Knowledge assistants

Knowledge assistants are AI copilots that answer questions and perform tasks using your documents—contracts, policies, procurement records, legal filings. Bundata supplies the document intelligence layer: context-aware extraction, smart bites in the Vector Catalog, and Vector Search so the assistant can retrieve relevant context and deliver grounded answers with source lineage.

What you need from Bundata

• Extraction — Documents turned into schema-aware chunks with metadata and optional enrichment (NER, table description). See Extraction overview.
• Vector Catalog — Indexed smart bites and embeddings. See Vector Catalog overview.
• Vector Search — Semantic retrieval so the assistant gets the right chunks for each user question. See Vector Search overview and Grounding from catalog.

Use cases

• Contract assistant — “What’s the termination clause in this agreement?” Query the catalog for the contract, retrieve relevant sections, and ground the answer with citations.
• Policy assistant — “What is our refund policy?” Search policy docs, return matching sections, and let the LLM summarize with source links.
• Legal / compliance — “Which contracts mention liability cap X?” Filter by metadata and search; return list with source lineage for review.
• Procurement / ops — “What did we agree with vendor Y?” Search contracts and invoices; ground answers in extracted terms and line items.

Design patterns

1. Index by use case — Use separate collections (or strong metadata) for contracts, policies, invoices so the assistant can scope queries. See Collections.
2. Retrieve then generate — Run Vector Search with the user’s question, pass top-k smart bites to the LLM, and generate an answer. Always attach source lineage to the response. See Grounding from catalog.
3. Filter when possible — Use metadata (date, document type, source) so retrieval is relevant. See Vector Search: Filtering.
4. Confidence and quality — Prefer high-extraction confidence bites; filter or flag low-confidence content so the assistant doesn’t ground on unreliable text. See Extraction best practices.

Common pitfalls

• Stale catalog — Keep ingestion updated with workflow orchestration so new and updated docs are searchable.
• No citations — Always show which document and (if possible) section each part of the answer came from.
• One-size-fits-all collection — Use multiple collections or strong metadata so the assistant doesn’t mix unrelated doc types.

Next steps

• Grounding from catalog — Retrieve and pass context to the LLM.
• First agent — Build your first agent.
• Vector Search best practices — Relevance and troubleshooting.