End of a long debugging rabbit hole. After ~10 attempts patching Graphiti's
compatibility with non-OpenAI LLMs (GLM, Qwen via SiliconFlow), every fix
uncovered another bug. Made the strategic call to abandon Graphiti entirely
for the graph-build path and write a minimal custom builder.
## Why Graphiti had to go
- Qwen 32B 32K context overflow on cumulative episode retrieval (60K-82K tokens)
- SiliconFlow Qwen 72B also 32K (not 128K as docs implied)
- GLM 4 Flash gives 128K context but triggers "20015 parameter invalid" on
Graphiti's structured output calls (logprobs in reranker, empty input in
embedder, nested dict in Neo4j writes)
- Each patch target was 1-2 levels deep in Graphiti internals
- 4 separate monkey-patches (EntityNode.save, bulk_utils, driver layer,
reranker, embedder) still couldn't cover the extract_nodes path
## New architecture
backend/app/services/custom_graph_builder.py (NEW, 348 lines):
- Uses Foresight's own LLMClient (retry + GLM→Qwen fallback + token tracker)
- 1 LLM call per chunk (Graphiti needed 4-5)
- Single prompt extracts entities + relationships as JSON
- Writes directly to Neo4j via cypher MERGE (no Graphiti dependency)
- Schema identical to what Graphiti produced — downstream get_all_nodes /
get_all_edges / zep_entity_reader all work unchanged
- ThreadPoolExecutor (10 workers) for concurrent chunk extraction
- Sequential Neo4j writes via shared session (avoids name conflicts in dedup)
- DDL wrapped in try/except (tolerates pre-existing Graphiti indexes)
- Entity name dedup via in-memory map → single uuid per canonical name
- Regex-sanitized label / relation type to prevent cypher injection
backend/app/api/graph.py:
- /api/graph/build endpoint now calls CustomGraphBuilder instead of
builder.add_text_batches() → client.add_episodes_batch() (the old Graphiti path)
backend/app/services/graph_builder.py:
- _build_graph_worker also updated to use CustomGraphBuilder (dead code path
but kept in sync)
backend/app/models/project.py:
- Default chunk_size 500 → 250 (to keep individual LLM prompts small)
backend/app/services/graphiti_client.py:
- Kept all monkey-patches for backward compat — they now only affect legacy
Graphiti code paths that CustomGraphBuilder bypasses entirely:
* _patch_neo4j_driver (AsyncSession.run nested-dict sanitize)
* _patch_reranker_for_non_openai (GLM logprobs workaround)
* _patch_embedder_empty_input (SiliconFlow empty-input guard)
* _patch_entity_node_ops (EntityNode.save sanitize)
* _patch_add_nodes_and_edges_bulk_tx (bulk-episode sanitize)
- These are kept for safety; Graphiti code path is no longer invoked in the
production build flow but methods like add_episode still exist on the class
## Performance
- Sequential: ~194 chunks × 2-5s = 10-15 min
- Concurrent (10 workers): ~194 / 10 × 3-5s = **1-2 min** (5-8x speedup)
- Rate limiting handled by LLMClient retry/backoff, not raw thread contention
## Downstream compatibility
Verified:
- get_all_nodes: MATCH (n:Entity) WHERE n.group_id = $gid RETURN n, labels(n)
- get_all_edges: MATCH (a)-[r]->(b) WHERE r.group_id = $gid RETURN r, type(r)
- get_node / get_node_edges: MATCH by uuid
CustomGraphBuilder writes:
- (n:Entity [optional second label]) with uuid, name, summary, group_id, created_at
- [r:RELATION_TYPE] with uuid, name, fact, group_id, created_at
Schema matches exactly.
## Pipeline wiring
Step 1 ontology generation → Step 2 graph build (CustomGraphBuilder) →
Step 3 profile generation (ZepEntityReader queries Neo4j) → Step 4 simulation
No changes needed downstream of Step 2.
7443082846