Data pool search settings

Semantic search, passages per answer, deep search and result order — what the four settings do and how to set them

Where to find the settings

Open Settings → Data sources, click the data pool and stay on the Overview tab. Four settings there decide how this pool is searched.

They apply per data pool, not per assistant — they take effect the same way whether the pool is used directly in a chat or through an assistant.

SettingDefaultWhat it does
Semantic searchonAdditionally finds what matches in meaning but is worded differently
Passages per answer10 (1–100)Upper limit of passages that go into one answer
Deep search (experimental)offHands retrieval to a dedicated search agent
Improve result orderoffCurrently has no effect — see below

Recommendation: leave it on. It is the reason search still works when your users phrase things differently from your documents. Someone asks "how do I bleed the valve", the manual says "venting the valve body" — with semantic search that is found, without it only the exact term is.

Turning it off is worth it only when this pool is searched by exact terms anyway — case numbers, part numbers, proper names — and the corpus is large. Indexing then finishes far sooner and costs less.

Note

Turning it off deletes nothing. It only stops new files from being embedded — whatever is already embedded stays searchable. Turning it back on does not take effect immediately: semantic search is rebuilt file by file from the next sync run.

Passages per answer

This value sets how many passages from this pool may go into a single answer. The range is 1 to 100, the default is 10.

Start at 10

The default fits most corpora. Only change it once you see a concrete symptom.

Go higher when answers are incomplete

Typical symptom: an answer breaks off mid-content, or a step is missing from a step-by-step instruction. This mostly happens with long operating manuals and handbooks containing tables. Raise the value in small steps and re-check with test search after each change.

Go lower for short, uniform documents

For FAQ collections or policies, fewer passages often produce the more precise answer, because fewer marginal hits come along.

Attention

Don't push this much higher than needed. More passages make the answer slower and take up room in the model's context window that is then missing for the conversation itself — pushed far enough, this can trigger a "chat too large" error. Raise it step by step and stop once test search confirms the answer is complete; there's no fixed number that's always safe. More on this: Chat too large for the model.

When deep search is active, a dedicated search agent takes over retrieval: it issues several queries in sequence, narrows or broadens them, and loads whole files when needed. For open-ended research tasks this can produce better answers, but it takes longer.

Three things to know before turning it on:

  • It has to be enabled twice. The switch on the data pool alone is not enough — deep search also has to be enabled for your organization. If it is not, the switch has no effect. Contact support@meingpt.com to get it enabled.
  • It exists for cloud data pools only. On an Outpost the platform deliberately refuses the search agent, because it cannot use the browse and full-text tools there. That is why the switch is not shown at all on Outpost data pools.
  • It replaces the normal retrieval path instead of adding to it. While it is active, the assistant only sees the search agent; the direct search and read tools are hidden.

Attention

Do not turn it on during an ongoing evaluation. Because deep search swaps out the entire retrieval path, answer behaviour changes noticeably. Deep-search results are then no longer comparable with results from the standard path. Evaluate the standard case first, then look at deep search deliberately on a single pool.

Improve result order

This setting currently has no effect

The switch stores a value that is not evaluated in any search path — neither for cloud data pools nor on the Outpost. Turning it on or off changes nothing about your search results.

Leave it off. If you want to influence the order of your hits, the effective levers are the scope of the data pool and semantic search, not this switch. We document this openly so that you do not attribute differences in answer quality to a setting that does nothing.

If you want to judge the quality of a data pool, test search tells you more than any of these four settings. You find it in the data pool on the Test Search tab.

There you put real end-user questions directly to the index. Every hit shows you which document and which passage it comes from, and you can switch between the search methods — keyword search, semantic search or both combined. That measures index quality directly instead of inferring it from chat answers.

An approach that works: collect five to ten questions your colleagues would actually ask and work through them in test search. If a question returns no fitting document, it is almost always the scope of the pool or a document that is not in the index — not one of the four settings. How to improve the scope: Narrow the data scope.

What is not configurable

Some search values are fixed and deliberately do not appear in the interface:

  • The mix between semantic search and keyword search in the combined search method.
  • The minimum relevance score below which a hit is discarded.
  • The choice of search method in chat. The platform decides that automatically per request. Only in test search can you switch the method by hand.

These values are tuned across corpora; making them individually adjustable would cause more misconfiguration than benefit.

Was this page helpful?