feat: add ENTITY_ATTRIBUTES support for domain-specific entity enrich…

[akshay-saraswat]

Allows operators to configure per-entity attribute extraction via the ENTITY_ATTRIBUTES environment variable (JSON list of attribute names), without any schema migrations or breaking changes to existing deployments.

Domain-specific LightRAG deployments often need entities enriched with context beyond name, type, and description. For example:

• A customer support knowledge graph benefits from sentiment and urgency entities.
• A research graph benefits from confidence scores on extracted claims.

Previously this required forking and hand-editing the extraction prompt with no supported mechanism to pass the attribute list through config.

The feature is implemented as a clean opt-in extension to the existing entity-extraction pipeline:

1. ENTITY_ATTRIBUTES env var (default: empty list) is parsed in config.py and passed through addon_params → extract_entities, the same path used by ENTITY_TYPES.

2. When non-empty, two placeholders are injected into the extraction prompts:

   • {entity_attributes}: the comma-separated attribute list
   • {entity_attributes_instruction}: a human-readable instruction string telling the LLM to append a compact single-line JSON object as a 5th field on each entity line.

3. _handle_single_entity_extraction now accepts 4 OR 5 fields. The 5th field is parsed as JSON → dict; malformed JSON is logged and silently dropped so extraction still succeeds for the entity itself.

4. Attributes are merged (last non-null value per key wins) across all extraction instances of the same entity, both in _merge_nodes_then_upsert (normal insert path) and _rebuild_single_entity (rebuild path).

5. The merged attributes dict is serialised as a JSON string and stored in the graph node under the key 'attributes'. All existing graph backends (NetworkX, Postgres, Neo4j, etc.) can round-trip it without schema changes. Consumers deserialise with json.loads.

Entity line with attributes:
entity<|#|>Pain Point X<|#|>painpoint<|#|>Users report...<|#|>{"sentiment":"negative","urgency":"high","confidence":0.87}

Stored node property:
"attributes": "{"sentiment": "negative", "urgency": "high", "confidence": 0.87}"

• ENTITY_ATTRIBUTES defaults to [] → extraction prompt is byte-for-byte identical to pre-patch behaviour.

• Parser still accepts 4-field records → existing extracted graphs are unaffected.

• 5-field records when ENTITY_ATTRIBUTES is empty are treated as malformed (entity still extracted without attributes).

  ENTITY_ATTRIBUTES='["sentiment","urgency","confidence"]'

Combined with ENTITY_TYPES for a customer support graph:

ENTITY_TYPES='["Person","Product","PainPoint"]'
ENTITY_ATTRIBUTES='["sentiment","urgency","confidence"]'

Description

• Adds ENTITY_ATTRIBUTES env var (JSON list of strings, default []) that requests extra per-entity attributes to be extracted alongside name/type/description.
• When non-empty, injects an instruction into the extraction prompt asking the LLM to append a compact single-line JSON object as a 5th field on each entity line.
• Parses and merges the 5th field across all extraction instances of the same entity; stored as a JSON string under node["attributes"].
• Zero change to default behaviour — empty ENTITY_ATTRIBUTES produces byte-identical prompts and 4-field records as before.

Related Issues

N/A

Changes Made

File	Change
lightrag/constants.py	DEFAULT_ENTITY_ATTRIBUTES = []
lightrag/api/config.py	Read ENTITY_ATTRIBUTES env var; pass into args.entity_attributes
lightrag/api/lightrag_server.py	Pass entity_attributes through addon_params
lightrag/prompt.py	Document optional 5th field in system prompt; add {entity_attributes_instruction} to user prompt
lightrag/operate.py	Read from addon_params; build instruction string; thread through _process_extraction_result and _handle_single_entity_extraction; merge and store attributes in both insert and rebuild paths

Checklist

• Changes tested locally
• Code reviewed
• Documentation updated (if necessary)
• Unit tests added (if applicable)

Additional Notes

N/A