🧠 Recall Plugin Design
The Recall Plugin performs a two-phase enrichment strategy grounded in contextual prompting and graph-based personalization for LLM agents.
1. Overview
The Recall Plugin helps LLM agents retrieve and enrich user prompts with relevant, personalized knowledge from a graph database. It operates in two distinct phases: intent detection and context enrichment.
2. Phase 1: Intent Detection & Scope Determination
- Input:
- User's natural language prompt
- A summary list of node types and relation types in the user's GraphDB (not full data)
- Action:
- Forwards the prompt and the list of node/relation names to the Handle (LLM-based orchestrator)
- The Handle infers the user's intent and responds with a directive:
- Which node types and relationships from the GraphDB are likely to help answer this prompt
- A suggested query scope, not a raw answer
- If the graph is empty or lacks relevant nodes, the Handle acknowledges the absence and adjusts response expectations accordingly
Example:
| Prompt | Graph Summary | Handle Response |
|---|---|---|
| "I broke up with my lover. What should I do?" | ["school", "career", "preferences"] | "There is no relational data about the user's romantic partner, but general personal preferences exist. Generate advice based on personality and life history." |
3. Phase 2: Enriched Prompt Completion
- Input:
- Original prompt
- Queried data from the relevant subset of the user's GraphDB (as determined in Phase 1)
- Action:
- The Recall Plugin enriches the prompt with this scoped context before passing it to the LLM for completion
4. Summary
The Recall Plugin enables:
- Personalized, context-aware LLM responses
- Efficient use of graph data by scoping queries to only relevant nodes/relations
- Graceful handling of missing or sparse user data