Knowledge
Your agent answers from documents you upload. This is what happens to one, and why the answer can still be wrong.
Why a document has to be cut up
A model cannot be handed your whole handbook for every question — it would be slow, expensive, and worse at finding the relevant line. Instead each document is split into passages, each passage is turned into a list of numbers that captures its meaning, and a question finds its answer by looking for the passages whose numbers are closest.
The actual numbers
| Passage size | 300 words |
|---|---|
| Overlap between passages | 50 words — so a sentence split across a boundary still appears whole in one of them |
| Passages used per answer | 5, the closest to the question |
| Embedding model | gemini-embedding-001 |
| Index | HNSW on cosine distance, in Postgres via pgvector |
| Formats | PDF, Word, plain text, Markdown, and a URL we fetch |
Why an answer can be wrong
Three different failures look identical to a visitor, and only one is the model's fault.
| The document was wrong | An agent repeats an out-of-date price as confidently as a current one. It has no way to know. This is the most common cause by far. |
|---|---|
| Retrieval missed | The answer exists, but the question was worded so differently from the passage that it was not among the five closest. Adding an FAQ phrased the way customers actually ask fixes more of these than any setting. |
| The model blended two passages | Two similar policies retrieved together can be merged into a third that matches neither. Rarer, and the reason a single authoritative document beats five overlapping ones. |
Deleting a document does not un-say it
If the agent keeps saying it does not know
What your plan limits
Plans cap the number of documents and the size of each file, not how many questions those documents can answer. A single thorough handbook is worth more than fifty thin pages, and costs one slot.
Next
Tools — letting an agent do things