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

# Retrieval

> Context Construction Pipeline

Retrieval in Persona is a process of context construction, not document fetching. The goal is to assemble a coherent picture of the user that informs the LLM's response.

## Tuning Parameters

The retrieval process is controlled by two parameters:

| Parameter   | Effect                                                                                                                                                                         |
| ----------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ |
| `top_k`     | Number of seed nodes from vector search. Increase for recall, decrease for precision.                                                                                          |
| `hop_depth` | Graph traversal depth from seeds. `0` = pure vector. `1` = immediate neighbors (previous episode, source psyche). `2+` = deeper causal chains. Higher depths increase latency. |

## Static Context

Before any query processing, the system fetches "always-on" context. This currently includes:

* **Active Goals**: All goals not marked `COMPLETED`.
* **Core Psyche**: The first 5 psyche traits (traits, preferences, values).

This ensures the agent always knows the user's current objectives and baseline identity, even if the query doesn't mention them.

## Why Graph Traversal Matters

Consider the query: "What happened before my meeting with Sarah?"

Vector search can find documents mentioning "Sarah" and "meeting." But it cannot answer "before." Only graph traversal can follow the `PREVIOUS` edge to return the chronologically preceding episode.

Similarly, a query about "why I'm stressed" might match an episode about stress. Graph traversal finds the `derived_from` link connecting that episode to a goal with an approaching deadline—the actual cause.

This is the core value of the hybrid approach: vector finds the topic, graph finds the context.

## Output Format

The final context is returned as structured XML to the LLM:

```xml theme={null}
<memory_context>
  <goals>
    <task status="active">Finish Auth Module</task>
  </goals>
  <psyche>
    <trait>Prefers morning work sessions</trait>
  </psyche>
  <episodes>
    <episode date="Yesterday">
      I'm struggling with the Auth0 integration...
    </episode>
    <episode date="Two Days Ago">
      Decided to switch from Firebase to Auth0.
    </episode>
  </episodes>
</memory_context>
```

This format groups memories by type, giving the LLM a structured view of the user's state, identity, and recent history.