botserver/docs/src/chapter-03/semantic-search.md

# Semantic Search

Semantic search enables the bot to retrieve information based on meaning rather than exact keyword matches. It leverages the vector embeddings stored in Qdrant.

## How It Works

1. **Query Embedding** – The user’s query string is converted into a dense vector using the same embedding model as the documents.
2. **Nearest‑Neighbor Search** – Qdrant returns the top‑k vectors that are closest to the query vector.
3. **Result Formatting** – The matching document chunks are concatenated and passed to the LLM as context for the final response.

## Using the `FIND` Keyword

```basic
SET_KB "company-policies"
FIND "how many vacation days do I have?" INTO RESULT
TALK RESULT
```

- `SET_KB` selects the collection.
- `FIND` performs the semantic search.
- `RESULT` receives the best matching snippet.

## Parameters

- **k** – Number of results to return (default 3). Can be overridden with `FIND "query" LIMIT 5 INTO RESULT`.
- **filter** – Optional metadata filter, e.g., `FILTER source="policy.pdf"`.

## Best Practices

- Keep the query concise (1‑2 sentences) for optimal embedding quality.
- Use `FORMAT` to clean up the result before sending to the user.
- Combine with `GET_BOT_MEMORY` to store frequently accessed answers.

## Performance

Semantic search latency is typically < 100 ms for collections under 50 k vectors. Larger collections may require tuning Qdrant’s HNSW parameters.
-												Add comprehensive documentation for GeneralBots, including keyword references, templates, and user guides

- Created detailed markdown files for keywords such as HEAR, TALK, and SET_USER.
- Added examples and usage notes for each keyword to enhance user understanding.
- Developed templates for common tasks like enrollment and authentication.
- Structured documentation into chapters covering various aspects of the GeneralBots platform, including gbapp, gbkb, and gbtheme.
- Introduced a glossary for key terms and concepts related to GeneralBots.
- Implemented a user-friendly table of contents for easy navigation.

											
										
										
											2025-10-25 14:50:14 -03:00
+								# Semantic Search
-												Revise documentation in Chapter 01 to improve clarity and structure, including updates to the installation instructions and session management overview.

											
										
										
											2025-10-25 15:59:06 -03:00
 								Semantic search enables the bot to retrieve information based on meaning rather than exact keyword matches. It leverages the vector embeddings stored in Qdrant.
 								## How It Works
 . **Query Embedding** – The user’s query string is converted into a dense vector using the same embedding model as the documents.
 . **Nearest‑Neighbor Search** – Qdrant returns the top‑k vectors that are closest to the query vector.
 . **Result Formatting** – The matching document chunks are concatenated and passed to the LLM as context for the final response.
 								## Using the `FIND` Keyword
 								```basic
 								SET_KB "company-policies"
 								FIND "how many vacation days do I have?" INTO RESULT
 								TALK RESULT
 								```
 								- `SET_KB` selects the collection.
 								- `FIND` performs the semantic search.
 								- `RESULT` receives the best matching snippet.
 								## Parameters
 								- **k** – Number of results to return (default 3). Can be overridden with `FIND "query" LIMIT 5 INTO RESULT`.
 								- **filter** – Optional metadata filter, e.g., `FILTER source="policy.pdf"`.
 								## Best Practices
 								- Keep the query concise (1‑2 sentences) for optimal embedding quality.
 								- Use `FORMAT` to clean up the result before sending to the user.
 								- Combine with `GET_BOT_MEMORY` to store frequently accessed answers.
 								## Performance
 								Semantic search latency is typically < 100 ms for collections under 50 k vectors. Larger collections may require tuning Qdrant’s HNSW parameters.