Skip to main content
An agent is only as good as the context it retrieves. If search returns plausible-but-wrong records, the agent reasons confidently over the wrong facts — and no amount of prompting fixes bad retrieval. This page is about making the agent’s own search accurate enough to trust. The common pattern is: your system runs RAG or prompt expansion, picks some chunks, and injects them into the prompt. This has three problems:
  • You guess wrong. Your retrieval heuristic decides what’s relevant before the agent knows what it needs for this specific step.
  • It poisons. Injected chunks the agent didn’t ask for are exactly the surface for context poisoning.
  • It’s opaque. Neither you nor the user can tell what actually drove the answer.
The better model: give the agent a search tool plus navigational meta-information, and make discovery its responsibility. It searches, reads what it got, decides whether to search again or drill into a relationship — the same way a good analyst works.
This is agentic search over all your structured content, not a bolt-on RAG index. In Open Index the agent gets navigation_guidelines() (what doc_types, fields, and relationships exist), search_brain(), and get_entity() — enough to navigate a domain it’s never seen.

​
Give the agent a map before the first query

Retrieval accuracy jumps when the agent knows the shape of what it’s searching. Before its first query, hand it navigational context:
  • Which doc_types exist (service, runbook, incident…) and what each means.
  • Which fields each carries, and which are searchable.
  • Which relationships connect them, so it can traverse on purpose.
With that map, the agent scopes searches (search only incidents), follows edges deliberately, and stops guessing. Without it, every query is a blind shot.

​
Tune ranking so the right record wins

Precision is also a ranking problem. Two levers matter most:
Weight a match in name far above a match in a long description. A genuine multiplier — a hit in a boost: 6 title outranks a boost: 1 body hit 6-to-1 — so the record whose identity matches the query surfaces first. → Search configuration
Keyword search nails exact names and identifiers; semantic search rescues queries that describe a thing in different words than the record uses (“checkout is slow” → “latency spike at payment”). Blend them — keyword-dominant by default, with a semantic weight you raise when users describe rather than name.

​
Design schemas for retrieval

Search accuracy starts at the schema, not the query:
  • Mark identity fields (name, owner, status) syntactic — exact/prefix matching, high boost.
  • Mark prose fields (description, summary) semantic — matched by meaning.
  • Mark opaque blobs and internal ids none — stored but never searched, so they can’t add noise to results.
A field that shouldn’t influence ranking simply isn’t indexed. That’s the cheapest precision win there is.

​
Checklist

  • The agent searches; you don’t pre-inject a guessed context blob.
  • It gets navigational meta-info (doc_types, fields, relationships) up front.
  • Identity fields are boosted; noise fields are search: none.
  • Ranking blends keyword + semantic to match both names and descriptions.

Next: Conflicts & staleness

Keeping retrieved context correct as it grows.