Problem

The zeph_graph_entities Qdrant collection (121 points) contains low-value structural tokens extracted from code and config rather than meaningful semantic entities about the user, their projects, and domain knowledge.

Observed examples (actual data from collection)

Extracted entities include:

• type, provider_type, allowed_commands, recall_limit — TOML config keys
• https, orchestrator, vector_collections — generic terms
• go — programming language with no context
• src/ — file path fragment
• read_file, wget — tool names

Expected behavior

Graph entities should represent:

• People: user identity, collaborators
• Projects: repo names, purpose, tech stack
• Decisions: architectural choices, preferences
• Domain concepts: meaningful nouns from conversations

Root cause hypothesis

The LLM-based entity extractor in graph_commands.rs / GraphMemory is:

1. Applied to tool outputs (code, config files) in addition to conversational messages, or
2. Using a prompt that does not filter structural/syntactic tokens, or
3. Applying extraction to raw JSON/TOML content that floods the entity graph with keys

Impact

• Entity graph is polluted — semantic search returns noise
• Graph-based recall (fetch_graph_facts) retrieves irrelevant context
• BFS traversal from meaningful nodes hits dead ends through noise nodes
• 121 noisy entities degrade rather than improve response quality

Suggested fix

1. Limit entity extraction to conversational messages only (exclude tool results)
2. Add entity type filter to extraction prompt: only extract Person, Project, Technology, Decision, Preference, Location
3. Add minimum entity length filter (reject single words like "go", "type")
4. Add deduplication pass with EntityResolver before upsert