Enhanced the edge deduplication prompts to better recognize semantically
equivalent facts that use different phrasings:
- Self-referential relationships ("X is a sub-agency of X" = "X is its own sub-agency")
- Active vs passive voice ("A awarded contract to B" = "B received contract from A")
- Numeric format equivalence ($1M = $1,000,000)
- Entity aliases (DoD = Department of Defense)
Added integration tests that verify the LLM correctly identifies semantic
duplicates with the improved prompts.
Changes to `to_prompt_json()` helper to default to minified JSON (no indentation) instead of 2-space indentation. This reduces token consumption in LLM prompts while maintaining all necessary information.
- Changed default `indent` parameter from `2` to `None` in `prompt_helpers.py`
- Updated all prompt modules to remove explicit `indent=2` arguments
- Minor code formatting fixes in LLM clients
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-authored-by: Claude <noreply@anthropic.com>
* Refactor prompt structure: move MESSAGES after instructions
Reordered prompt structure in extract_nodes.py to place MESSAGES section
after instructions/guidelines in both extract_attributes and extract_summary
functions for improved prompt clarity.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
* Add sentence-aware text truncator for entity summaries
- Created truncate_at_sentence() utility function that truncates text at
sentence boundaries while respecting max character limits
- Added MAX_SUMMARY_CHARS constant (250 chars) for entity summaries
- Applied truncator to entity summaries in prompts (extract_nodes.py)
- Applied truncator to LLM-generated summaries (node_operations.py)
- Added comprehensive test suite for truncation logic
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
* Clean up formatting in extract_attributes prompt
- Remove extra blank lines
- Fix indentation of MESSAGES tag
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
* Bump version to 0.22.0pre3
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
---------
Co-authored-by: Claude <noreply@anthropic.com>
* Refactor summary prompts to use character limit and prevent meta-commentary
- Changed summary length constraint from "8 sentences" to "250 characters" for more predictable output
- Created reusable summary_instructions snippet in snippets.py with clear BAD/GOOD examples
- Added explicit instruction to output only factual content without meta-commentary
- Applied consistent formatting across extract_nodes.py and summarize_nodes.py
- Bumped version to 0.22.0pre2
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
* Add copyright header to snippets.py
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
---------
Co-authored-by: Claude <noreply@anthropic.com>
* Enforce shorter summaries with 8 sentence limit
Replace 250-word limit with 8 sentence limit for node summaries to improve conciseness. Also update prompt system message for summarize_context to better reflect its dual purpose of generating summaries and attributes.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
* Update graphiti_core/prompts/summarize_nodes.py
Co-authored-by: claude[bot] <209825114+claude[bot]@users.noreply.github.com>
* Bump version to 0.22.0pre1
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
* Update graphiti_core/prompts/summarize_nodes.py
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
---------
Co-authored-by: Claude <noreply@anthropic.com>
Co-authored-by: claude[bot] <209825114+claude[bot]@users.noreply.github.com>
* fix: Improve edge extraction entity ID validation
Fixes invalid entity ID references in edge extraction that caused warnings like:
"WARNING: source or target node not filled WILL_FIND. source_node_uuid: 23 and target_node_uuid: 3"
Changes:
- Format ENTITIES list as proper JSON in prompt for better LLM parsing
- Clarify field descriptions to reference entity id from ENTITIES list
- Add explicit entity ID validation as #1 extraction rule with examples
- Improve error logging (removed PII, added entity count and valid range)
These changes follow patterns from extract_nodes.py and dedupe_nodes.py where
entity referencing works reliably.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
* wip
* fix: Align fact field naming and add description
- Change extraction rule to reference 'fact' instead of 'fact_text'
- Add descriptive text for fact field in Edge model
* fix: Remove ensure_ascii parameter from to_prompt_json call
Align with other to_prompt_json calls that don't use ensure_ascii
* fix: Use validated target_node_idx variable consistently
Line 190 was using raw edge_data.target_entity_id instead of the
validated target_node_idx variable, creating inconsistency with line 189
* fix: Improve edge extraction validation checks
- Add explicit check for empty nodes list
- Use more explicit 0 <= idx comparison instead of -1 < idx
- Prevents nonsensical error message when no entities provided
* chore: Restore uv.lock from main branch
Previously deleted in commit 7e4464b, now restored to match main branch state
* Update uv.lock
---------
Co-authored-by: Claude <noreply@anthropic.com>
* Optimize edge deduplication prompt for caching and clarity
- Restructure prompt to place invariant instructions at top and dynamic context at bottom for better LLM caching
- Change 'id' to 'idx' in edge context lists to avoid confusion with other identifiers
- Remove 'fact_type_id' from edge types context as LLM only needs fact_type_name
- Remove dynamic range values from prompt instructions (e.g., "range 0-N")
- Add debug logging before LLM call to track input sizes
- Add validation logging after LLM response to catch invalid idx values
- Clarify that duplicate_facts uses EXISTING FACTS idx and contradicted_facts uses INVALIDATION CANDIDATES idx
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
* Address terminology consistency and edge case logging
- Update Pydantic field descriptions to use 'idx' instead of 'ids' for consistency
- Fix debug logging to handle empty list edge case (avoid 'idx 0--1' display)
Note on review feedback:
- Validation is intentionally non-redundant: warnings provide visibility, list comprehensions ensure robustness
- WARNING level is appropriate for LLM output issues (not system errors)
- Existing test coverage is sufficient for this defensive logging addition
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
---------
Co-authored-by: Claude <noreply@anthropic.com>
* Remove ensure_ascii configuration parameter
- Changed to_prompt_json default from ensure_ascii=True to False
- Removed ensure_ascii parameter from Graphiti.__init__ and GraphitiClients
- Removed ensure_ascii from all function signatures and context dictionaries
- Removed ensure_ascii from all test files
- All JSON serialization now preserves Unicode characters by default
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
* format
---------
Co-authored-by: Claude <noreply@anthropic.com>
* fix: Improve deduplication ID validation and logging
- Add comprehensive logging to verify IDs sent to LLM (sent vs received)
- Enhance prompt with explicit ID bounds (0 through N-1)
- Add validation warnings for missing and extra IDs from LLM responses
- Improve error message clarity for invalid dedupe IDs
- Log actual IDs sent to LLM to confirm no index leakage
This helps diagnose cases where the LLM returns IDs outside the valid
range (e.g., ID 19 when only 0-18 were sent).
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
* fix: Remove redundant logging parameter
Address reviewer comment about redundant third parameter in debug log statement.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
* fix: Address reviewer comments on list slicing and prompt clarity
- Fix list slicing bug: change <= to < to avoid gap when exactly 20 elements
(previously would skip element 10 when showing 21 elements)
- Consolidate redundant prompt phrasing while maintaining clarity
(reduced from 3 sentences to 2, keeping essential constraints)
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
* fix: Remove redundant prompt text to reduce token usage
Consolidate 'using these exact IDs (0 through N-1)' with following sentence
to eliminate repetition. Changes:
- 'using these exact IDs (0 through {N-1}). Do not skip IDs or use IDs outside this range'
- 'with IDs 0 through {N-1}. Do not skip or add IDs'
Saves ~15 tokens per deduplication call.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
---------
Co-authored-by: Claude <noreply@anthropic.com>
* fix: Fix typo in JSON entity extraction prompt
Change "an entities" to "any entities" in guideline 1 of the extract_json prompt.
🤖 Generated with [Claude Code](https://claude.ai/code)
Co-Authored-By: Claude <noreply@anthropic.com>
* Update graphiti_core/prompts/extract_nodes.py
Co-authored-by: Daniel Chalef <131175+danielchalef@users.noreply.github.com>
---------
Co-authored-by: Claude <noreply@anthropic.com>
Co-authored-by: Daniel Chalef <131175+danielchalef@users.noreply.github.com>
* chore: Update edge extraction prompt to paraphrase instead of quote
- Changed instruction 5 to request paraphrasing rather than verbatim quoting
- Updated string quotes to use double quotes for consistency
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
* chore: Format edge_operations.py and update lock file
- Minor formatting fix in edge_operations.py list comprehension
- Update uv.lock with version bump to 0.21.0rc8
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
---------
Co-authored-by: Claude <noreply@anthropic.com>
Add guideline to extract entities from all JSON properties, not just primary fields like name/user. This ensures comprehensive entity extraction while maintaining the existing exclusion of date properties.
🤖 Generated with [Claude Code](https://claude.ai/code)
Co-authored-by: Claude <noreply@anthropic.com>
* Add support for non-ASCII characters in LLM prompts
- Add ensure_ascii parameter to Graphiti class (default: True)
- Create to_prompt_json helper function for consistent JSON serialization
- Update all prompt files to use new helper function
- Preserve Korean/Japanese/Chinese characters when ensure_ascii=False
- Maintain backward compatibility with existing behavior
Fixes issue where non-ASCII characters were escaped as unicode sequences
in prompts, making them unreadable in LLM logs and potentially affecting
model understanding.
* Remove unused json imports after replacing with to_prompt_json helper
- Fix ruff lint errors (F401) for unused json imports
- All prompt files now use to_prompt_json helper instead of json.dumps
- Maintains clean code style and passes lint checks
* Fix ensure_ascii propagation to all LLM calls
- Add ensure_ascii parameter to maintenance operation functions that were missing it
- Update function signatures in node_operations, community_operations, temporal_operations, and edge_operations
- Ensure all llm_client.generate_response calls receive proper ensure_ascii context
- Fix hardcoded ensure_ascii: True values that prevented non-ASCII character preservation
- Maintain backward compatibility with default ensure_ascii=True
- Complete the fix for issue #804 ensuring Korean/Japanese/Chinese characters are properly handled in LLM prompts
* Bump version from 0.9.0 to 0.9.1 in pyproject.toml and update google-genai dependency to >=0.1.0
* Bump version from 0.9.1 to 0.9.2 in pyproject.toml
* Update google-genai dependency version to >=0.8.0 in pyproject.toml
* loc file
* Update pyproject.toml to version 0.9.3, restructure dependencies, and modify author format. Remove outdated Google API key note from README.md.
* upgrade poetry and ruff
* Update README.md to include installation instructions for Graphiti with Google Gemini support
* fix to deps since peotry doesn't fully implement PEP 735
* Refactor string formatting in various files to use single quotes for consistency and improve readability. This includes updates in agent.ipynb, quickstart.py, multiple prompt files, and ingest.py and retrieve.py modules.
* Remove optional dependencies from pyproject.toml to streamline project requirements.
* Bump version from 0.9.0 to 0.9.1 in pyproject.toml and update google-genai dependency to >=0.1.0
* Bump version from 0.9.1 to 0.9.2 in pyproject.toml
* Update google-genai dependency version to >=0.8.0 in pyproject.toml
* loc file
* Update pyproject.toml to version 0.9.3, restructure dependencies, and modify author format. Remove outdated Google API key note from README.md.
* upgrade poetry and ruff
* first cut
* Update dependencies and enhance README for optional LLM providers
- Bump aiohttp version from 3.11.14 to 3.11.16
- Update yarl version from 1.18.3 to 1.19.0
- Modify pyproject.toml to include optional extras for Anthropic, Groq, and Google Gemini
- Revise README.md to reflect new optional LLM provider installation instructions and clarify API key requirements
* Remove deprecated packages from poetry.lock and update content hash
- Removed cachetools, google-auth, google-genai, pyasn1, pyasn1-modules, rsa, and websockets from the lock file.
- Added new extras for anthropic, google-genai, and groq.
- Updated content hash to reflect changes.
* Refactor import paths for GeminiClient in README and __init__.py
- Updated import statement in README.md to reflect the new module structure for GeminiClient.
- Removed GeminiClient from the __all__ list in __init__.py as it is no longer directly imported.
* Refactor import paths for GeminiEmbedder in README and __init__.py
- Updated import statement in README.md to reflect the new module structure for GeminiEmbedder.
- Removed GeminiEmbedder and GeminiEmbedderConfig from the __all__ list in __init__.py as they are no longer directly imported.
* implement so
* bug fixes and typing
* inject schema for non-openai clients
* correct datetime format
* remove List keyword
* Refactor node_operations.py to use updated prompt_library functions
* update example
