# Find the right context

Combine relevance signals without confusing a match with a fact.

Source: https://docs.minds.sh/docs/concepts/retrieval



Retrieval selects useful context for a task. An exact key lookup, an embedding search, a graph query, and a memory search can each find something different.

## Choose the signal [#choose-the-signal]

| Signal                    | Useful when                                                              |
| ------------------------- | ------------------------------------------------------------------------ |
| Exact identifier          | You already know the record you want                                     |
| Structured filter         | A field, namespace, or time range should restrict results                |
| Vector similarity         | The wording differs but the meaning or learned representation is related |
| Graph relationship        | The answer depends on how entities are connected                         |
| Memory-specific retrieval | The role of the information matters, such as an event or a procedure     |

Hybrid retrieval combines signals rather than forcing every question through one representation. Use the request's supported fields and weights; do not assume a UI search silently applies every possible signal.

## Similarity is not confidence [#similarity-is-not-confidence]

A high similarity score means the query and result are close under a particular representation and metric. It does not establish factual accuracy, recency, permission, or trust. Always apply access scope and inspect sources before turning retrieved text into an answer.

For vector indexes, keep embedding dimensions and model identity consistent. A change of embedding model can require a new index or re-embedding; mixing representations can produce misleading results even when requests succeed.

[Memory retrieval API](/docs/api/instance/memory) · [Vector API](/docs/api/instance/vectors) · [Search in the workspace](/docs/workspace/search)