recall: Graph expansion follows too many hops through generic/hub nodes

[jack-arturo]

Problem

Graph expansion during recall follows too many hops through generic entity nodes, pulling in memories that are connected only through shared infrastructure or common contacts — not through meaningful semantic relationships.

Example Traversal Path

Query: "Alex Panagis" with expand_entities=true

Alex Panagis (memory) 
  → entity:organizations:automem 
    → Alex Beck memories (also uses AutoMem)
      → entity:people:jack 
        → Zack Katz, Luka, Mastermind crew, every Jack-related memory

The expansion followed: person → tool → different person → shared contact → everyone. By hop 3, results have zero relevance to the original query.

Current Mitigation

The expand_min_importance and expand_min_strength params help somewhat (cut results from 40 → 16), but they filter on node properties, not on path relevance. A high-importance memory about Alex Beck is still irrelevant to an Alex Panagis query, regardless of its importance score.

Proposed Solutions

1. Hop-depth limiting by query type: For person/entity lookups, cap expansion at 1 hop. For topic/concept queries, allow 2-3 hops. The system could infer query type from the presence of proper nouns vs. general terms.

2. Path relevance decay: Apply a decay multiplier at each hop. If the seed result scores 0.72, the first hop should require at least ~0.5 relevance to the original query (not just to the intermediate node), second hop ~0.35, etc.

3. Hub node detection: Identify high-connectivity "hub" nodes (like entity:organizations:automem, entity:people:jack) that connect to many unrelated memories, and deprioritize expansion through them. These nodes are structurally important but semantically promiscuous.

4. Entity-type-aware expansion: When the query is about a person, only expand through entity:people:* nodes, not through entity:organizations:* or entity:tools:*. This prevents the tool/org bridge problem.

5. Configurable max_hops parameter: Let callers explicitly set maximum graph traversal depth (default: 2, person queries: 1).

Related Issues

• recall: Full-name entity disambiguation (Alex Panagis vs Alex Beck problem) #71 (entity disambiguation) — overlapping root cause
• enrichment: Entity extraction produces misclassified types and duplicate entities #72 (entity extraction quality) — bad entities create bad expansion paths
• recall: Add minimum relevance score threshold to filter low-quality results #73 (score floor) — expansion noise could be partially mitigated by score filtering

Impact

Uncontrolled expansion is the primary reason recall returns 40 results when 5 would suffice. Each irrelevant result costs context tokens and degrades the LLM's ability to synthesize a useful response.

[coderabbitai]

🔗 Related PRs

#8 - feat: Multi-hop recall with graph expansion and multi-query support [merged]

---

📝 Issue Planner

_{Check the box below or use the @coderabbitai plan command to generate an implementation plan and prompts that you can use with your favorite coding assistant.}

• Create Plan

---

🧪 Issue enrichment is currently in open beta.

You can configure auto-planning by selecting labels in the issue_enrichment configuration.

To disable automatic issue enrichment, add the following to your .coderabbit.yaml:

issue_enrichment:
  auto_enrich:
    enabled: false

💬 Have feedback or questions? Drop into our discord!

[WebsationFlow]

If you're dealing with graph expansion through generic nodes, you might find ConversationalFilter useful for managing scope creep in your LLM interactions. It helps keep responses focused by truncating at natural boundaries and adding clarifying questions. Check it out at github.com/WebsationFlow/conversational-filter.

[jack-arturo]

@WebsationFlow The quality of the actual LLM response hasn't been affected, at least in my experience. The problem is, for example... if I ask, "What do we know about Tim Cook?", I'd expect a graph traversal like:

Tim Cook -> Apple Corp -> Current iPhone model

And say, 8 memories (this is supposing Tim and I are buddies and AutoMem collects a lot of memories about our interactions 😅)

What we're getting at the moment does include that path, the responses are relevant. But also returns results like:

Tim Cook -> Apple Corp -> Apples -> Trees

Or

Tim Cook -> Tim Allen -> Toy Story

And can creep up to 40+ memories.

The LLM knows to discard the low relevant memories from the response and our interaction isn't affected, but it means slower recall due to the deeper traversal, and wasted tokens.