forward-port graphiti_core improvements from internal fork

[prasmussen15]

Summary

Forward-ports a set of graphiti_core improvements that have been living in an internal fork back to this repo. No workflow / CONTRIBUTING / CLA changes are included — those have moved forward here and we don't want to regress them.

Multi-episode batched extraction

Graphiti._extract_and_resolve_nodes and extract_edges now accept a list of episodes instead of a single one. When multiple are passed:

• Content is concatenated with [Episode N] headers via the new concatenate_episodes helper in graphiti_core/utils/text_utils.py.
• Each extracted node and edge carries episode_indices: list[int] so callers can map results back to the originating episode.
• Prompts (extract_nodes.extract_message, extract_edges.edge) and Pydantic models updated accordingly.
• build_episodic_edges (in edge_operations.py) accepts a single or multi episode UUID and a node_episode_index_map to wire up MENTIONED_IN edges per episode.

Files: graphiti_core/graphiti.py, graphiti_core/utils/text_utils.py, graphiti_core/utils/maintenance/{edge,node}_operations.py, graphiti_core/utils/bulk_utils.py, graphiti_core/prompts/{extract_edges,extract_nodes}.py, tests/utils/test_concatenate_episodes.py, tests/utils/maintenance/{test_bulk_utils,test_entity_extraction}.py, tests/{helpers_test,test_graphiti_mock}.py.

fact_triple episode type and episode_metadata

• New EpisodeType.fact_triple variant in graphiti_core/nodes.py.
• New episode_metadata: dict[str, Any] | None field on EpisodicNode for customer-defined filtering keys.

Safer attribute merging

When persisting an edge or node, the attributes dict was previously merged into edge_data / entity_data with dict.update(...), which would silently overwrite first-class fields (uuid, group_id, created_at, timestamps, etc.) if an attribute dict happened to contain a matching key. Now merged with setdefault semantics — existing typed fields win. Files: graphiti_core/edges.py, graphiti_core/utils/bulk_utils.py.

reference_time preservation in bulk edge save

reference_time is now passed through add_nodes_and_edges_bulk_tx alongside the other edge timestamps. File: graphiti_core/utils/bulk_utils.py.

Fact-extraction prompt refinements

Based on evaluation findings that concrete details (brand names, quantities, colors, qualified objects) were being lost during extraction:

• extract_edges: relaxed strict two-entity rule (look for a list-anchor before dropping), strengthened detail-preservation rule, clarified intra-CURRENT_MESSAGE dedup scope.
• extract_nodes: softened the "Wikipedia article" specificity test so brand-named and descriptor-qualified items ("Gamecube", "wool coat", "cracked windshield") are extracted while bare head nouns ("coat", "windshield") are not.

Files: graphiti_core/prompts/{extract_edges,extract_nodes}.py.

summarize_sagas prompt refresh

Updated prompt for the saga-summary path. File: graphiti_core/prompts/summarize_sagas.py.

Minor driver tweaks

graphiti_core/driver/graph_operations/graph_operations.py, graphiti_core/driver/record_parsers.py.

pytest bump

server/pyproject.toml and mcp_server/pyproject.toml bumped pytest>=9.0.3 with matching uv.lock refreshes.

Test plan

• make check (format + lint + tests) passes locally
• Existing unit tests pass (multi-episode code paths are exercised in the new test_concatenate_episodes.py and the updated test_entity_extraction.py / test_bulk_utils.py / test_graphiti_mock.py)
• Integration tests (pytest -k _int) — run locally with Neo4j before merge
• Reviewer: sanity check the episode_indices / node_episode_index_map plumbing in edge_operations.py and node_operations.py
• Reviewer: confirm the attribute-merging change doesn't break any custom edge/entity types that rely on the previous overwrite-on-collision behavior

🤖 Generated with Claude Code