- Increase chunk_top_k from 10 to 20
- Reduce max_entity_tokens to 6000
- Reduce max_relation_tokens to 8000
- Update web UI default values
- Fix max_total_tokens to 30000
Replace regex-based JSON extraction with json-repair for better handling of malformed LLM responses. Remove deprecated JSON parsing utilities and clean up keyword_extraction parameter across LLM providers.
- Remove locate_json_string_body_from_string() and convert_response_to_json()
- Use json-repair.loads() in extract_keywords_only() for robust parsing
- Clean up LLM interfaces and remove unused parameters
- Add json-repair dependency
This parameter is no longer used. Its removal simplifies the API and clarifies that token length management is handled by upstream text chunking logic rather than the embedding wrapper.
This commit renames the parameter 'llm_model_max_token_size' to 'summary_max_tokens' for better clarity, as it specifically controls the token limit for entity relation summaries.
- Remove MAX_TOKEN_SUMMARY parameter and related configurations
- Eliminate forced token-based truncation in entity/relationship descriptions
- Switch to fragment-count based summarization logic using FORCE_LLM_SUMMARY_ON_MERGE
- Update FORCE_LLM_SUMMARY_ON_MERGE default from 6 to 4 for better summarization
- Clean up documentation, environment examples, and API display code
- Preserve backward compatibility by graceful parameter removal
This change resolves issues where LLMs were forcibly truncating entity relationship
descriptions mid-sentence, leading to incomplete and potentially inaccurate knowledge
graph content. The new approach allows LLMs to generate complete descriptions while
still providing summarization when multiple fragments need to be merged.
Breaking Change: None - parameter removal is backward compatible
Fixes: Entity relationship description truncation issues
- Refactor the trigger condition for LLM-based summarization of entities and relations. Instead of relying on character length, the summary is now triggered when the number of merged description fragments exceeds a configured threshold. This provides a more robust and logical condition for consolidation.
- Introduce the `OLLAMA_NUM_CTX` environment variable to explicitly configure the context window size (`num_ctx`) for Ollama models. This decouples the model's context length from the `MAX_TOKENS` parameter, which is now specifically used to limit input for summary generation, making the configuration clearer and more flexible.
- Updated `README` files, `env.example`, and default values to reflect these changes.
- Add prominent warning about required initialization steps
- Document common errors (AttributeError: __aenter__ and KeyError: 'history_messages')
- Add troubleshooting section with specific solutions
- Add inline comments in code example highlighting initialization requirements
This addresses user confusion when LightRAG fails with cryptic errors due to
missing initialization calls. The documentation now clearly states that both
await rag.initialize_storages() and await initialize_pipeline_status() must
be called after creating a LightRAG instance.
