Navidev KB RAG

Search

Find the passages in your documents that talk about what you asked.

Search retrieves passages from your documents; it does not write answers. You type what you are looking for and the platform returns the pieces of text closest in meaning, telling you which document each one comes from and how good the match is.

If what you want is a written answer, that is the Ask screen.

Search or Ask

Both query your documents, but they serve different purposes:

SearchAsk
ReturnsLiteral passagesA written answer
Use it toRead the source in contextUnderstand something quickly
TraceabilityThe text and its scoreCites the documents used
CostLow: embeds your queryHigh: also writes the answer

A practical way to use them: Ask to get the gist quickly, Search when you need to read the literal source — before making a decision or quoting something, for instance.

What a search costs

Searching is not free, even though it costs little. To compare by meaning, the platform has to turn your query into a vector, and that conversion is a call to the knowledge base's embedding model, using your provider credential.

What you pay for is only the text you type, a few lines. The documents are not paid for again: that happened once, when they were indexed.

That is why a search costs considerably less than a question in Ask, where on top of embedding the query the answer has to be generated with a text model — far more expensive per token, and with many more tokens involved.

Pick the knowledge base and type what you are after. Enter runs the search and Shift + Enter inserts a line break.

You do not need to type isolated keywords or guess the document's exact wording: the search works by meaning. A natural phrase describing what you want usually works better than two loose words.

Reading the results

Results come grouped by document and ordered by relevance: documents with the best match first, and within each one its best passages.

Each document shows its three best passages. If it has more, a link reveals the rest. This is deliberate: a single document can hold fifteen matches and would bury every other one.

The file name links to its knowledge base's document list, in case you want to go to the full document.

The score

Every passage carries a score between 0 and 1: how close it is in meaning to your query.

ScoreColourHow to read it
0.70 or aboveGreenClear match
Between 0.50 and 0.70AmberRelated, read it with judgement
Below 0.50GreyWeak match

The text may look messy

Passages are shown exactly as they came out when the text was extracted from the original file. It is normal to find sentences cut mid-way, hyphenation or odd line breaks, especially in PDFs. That is not a bug: it is the text the platform actually indexed.

Downloading the original

When a document has its original file stored, a download action appears in the group's header.

If you see "Original not available" instead, that document was uploaded without keeping the original (see Knowledge bases). It is a property of each document, not of the base: a single search can mix documents with and without a stored original.

If you find nothing

When a search returns no results, try to:

  • Rephrase it. Search compares meanings, so a different turn of phrase may land closer to how the document is written.
  • Describe the idea in a sentence rather than using two or three loose terms.
  • Check you are in the right base and that its documents are in the Ready state. A pending or failed document is not searchable yet.

Request limit

Searches are limited to 20 per minute. If you run many in a row you will see a notice asking you to wait a few seconds.

This is a limit of this application, not of your account with the model provider: it has nothing to do with your credit or your key.

Updated on Aug 05, 2026