The answer can only be as good as the search
A colleague wants the current procedure for a service handover. The collection contains the current guide, an archived guide with an almost identical title, a transition checklist, and a meeting note that mentions the guide without changing it. Asking AI to summarize the first few matches is an easy way to produce a polished answer to the wrong question.
Before generating a sentence, make two choices visible. Which words will find the material? Which rules decide whether that material belongs in the answer? The small search bench below lets you change those choices and inspect what happens.
Read it.
Then explain it.
Find the current handover procedure.
Change the vocabulary or the version rule. Open a match to see what it supports.
handovercurrent records only| Matching record | Version | Content |
|---|---|---|
| Current | Procedure | |
| Current | Context |
Service handover guide
For a service handover, name the incoming owner, list the open actions, and ask the receiving team to confirm that it can access the working documents.
This passage gives a current procedure. It supports an instruction to confirm ownership, open actions, and document access.
The current guide supports the procedure. The meeting note is a match, but it does not introduce a new instruction.
This browser example uses literal word matching on an invented collection. It makes inclusion decisions visible; it is not a live database search or a quality benchmark.
A word and a rule do different work
Searching for “handover” finds records containing that word. Requiring a current record excludes an archived guide even when its wording is an excellent match. These decisions should remain separate. If the system quietly drops the current-only requirement to obtain more results, it has changed the question.
Broader vocabulary helps in a different way. A useful document may call the process a transition. Adding that term can find it without changing the requirement that the source be current. The reader should be able to see why an additional record appeared, rather than being asked to trust a mysterious relevance score.
Turn the current-only filter off and the older guide returns. Broaden the wording and a differently titled checklist becomes available.
The matching passage still needs reading
A meeting note can contain the right word without containing the procedure. It may say that someone needs to write the guide, or that a previous handover went badly. That is useful context, but weak support for an instruction.
Open a result and look at the excerpt before accepting the generated explanation. Ask what the passage actually establishes. A record can be relevant to the topic while leaving the specific question unanswered. This is why retrieval and assessment need separate checks, even in a small assistant.
Keep a stable identifier with each result. The eventual answer should point to the record that was read, not a title reconstructed from memory. Similar titles and multiple versions are ordinary features of business documents. They should not turn a source link into guesswork.
Use the search tools your data already supports
When records already live in PostgreSQL, its full-text search is a reasonable place to begin. PostgreSQL’s documentation describes full-text search that normalizes words, supports dictionaries and ranking, and can use indexes. Structured fields can handle requirements such as status, account, or date alongside the text query.
Inspect a set of failed searches before adding another retrieval system. If users and documents use different terms, a vocabulary change may help. If records are stale, fix ingestion. If a filter is too restrictive, clarify the requirement. A vector index should answer a demonstrated retrieval need, with an evaluation showing what improved and what regressed.
The agent’s database access also needs a boundary. A narrow tool can accept approved filters and search terms while the backend enforces access, query limits, and returned fields. If generated SQL is necessary, use restricted database privileges and execution limits. An instruction to read only is not a database permission.
Ask what the search left out
Good-looking results tell you something about precision. They tell you much less about relevant records that never appeared. Create a reviewed set of representative questions with expected evidence, including alternative wording and awkward version histories. When retrieval changes, inspect the useful records it gains and the ones it loses.
Coverage needs plain language. If a query returns a capped pool and only some of that pool is read, the answer describes the assessed subset. A fluent explanation cannot turn it into an exhaustive search. Show the limits where they affect the user’s next decision.
For our handover question, the next step is now concrete. Read the current guide, inspect the transition checklist if broader wording is enabled, and keep the meeting note in its proper role as context. Then write the explanation with those sources attached.