Test retrieval and settings

Ask a base a question the way a caller would, and tune the match cutoff, passages per answer, hybrid search and refresh.

What you will achieve

You will confirm a base finds the right passages, and tune how strict it is.

Step-by-step: test retrieval

Open the base and the Test retrieval tab.
Type a question as a caller would, for example "How long do I have to return something?"
Choose Search. This runs the same lookup as a live call and shows the passages found, with their scores and source documents.

If nothing comes back, either the document does not cover it or the Match cutoff is too strict.

Settings

Open the Settings tab. Changes save when you leave a field.

Setting	What it does
Embedded with	The model and dimensions used. Fixed once documents are indexed. Also shows the model's price and what the base has cost so far.
Match cutoff	How similar a passage must be to count. Higher means fewer, closer matches. The page shows the value that suits the model (about 0.30 for OpenAI's newer models, 0.65 for Gemini).
Passages per answer	How many passages go to the model (1 to 20, default 4). More passages cost more tokens every turn.
Hybrid search	Adds keyword matching to meaning. A passage containing a code, model number, price or phone number the caller said (such as "X-460") is always returned and listed first, even below the Match cutoff. Works in every language, including Urdu and Arabic. Leave on.
Automatic refresh	Hours between re-fetching web pages. 0 turns it off.

On each turn of a call, the agent's attached bases are searched with what the caller said. The best passages (up to the setting above) are given to the model, which answers from them. Attached bases are searched on every turn, so attach only what the agent needs.

Testing

Ask five real questions, including one the documents do not answer. A good base returns relevant passages for the first four and nothing misleading for the last.

Expected result

Each answerable question returns the passage that contains the answer near the top.

Troubleshooting

No results. Lower Match cutoff; check the document is Indexed.
Irrelevant results. Raise Match cutoff; split a huge document into focused ones.
Exact codes are missed. Turn Hybrid search on.
A follow-up question gets the wrong answer. Short follow-ups such as "How much does it cost?" are searched together with the caller's previous question. Greetings and "thanks" are not searched at all, so they add no delay.

Something here does not match what you see in the dashboard? Let us know.