> ## Documentation Index
> Fetch the complete documentation index at: https://docs.jabrod.com/llms.txt
> Use this file to discover all available pages before exploring further.

# Querying

> How retrieval works

# Querying & Retrieval

Once your data is chunked, embedded, and stored, you can query it via the `/api/rag/query` endpoint.

## The Retrieval Process

When you send a query:

1. **Query Embedding**: Jabrod converts your query string into a vector using the pipeline's configured embedding model.
2. **Vector Search**: Jabrod searches the vector database for chunks whose vectors are closest (most similar) to the query vector.
3. **Filtering**: Any results below the pipeline's configured **Similarity Threshold** are discarded.
4. **Ranking**: The top `K` results are returned to you, ordered by similarity score.

## Top K

The `topK` parameter determines how many chunks are returned.

* A low `topK` (e.g., 3) returns only the most highly relevant pieces of information, which is cheaper and faster if you are feeding it into an LLM.
* A high `topK` (e.g., 10-20) returns more context, which is useful if the answer is spread across multiple documents, but increases token usage if passed to an LLM.

You configure a default `retrievalTopK` on the pipeline itself, but you can override it on a per-request basis in the API.

## Similarity Threshold

Vectors are compared using Cosine Similarity, which results in a score between 0 and 1.

* `1.0` means identical meaning.
* `0.0` means completely unrelated.

If you set a **Similarity Threshold** of `0.7`, Jabrod will ignore any chunks that score below `0.7`. This helps prevent "hallucinations" by ensuring that if no relevant data exists in the knowledge base, the API returns an empty array rather than returning vaguely related garbage.