RomanNum3ral
/
MicroFish
mirror of https://github.com/666ghj/MiroFish.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity
272 Commits 1 Branch 3 Tags 74 MiB
Commit Graph

6 Commits

Author SHA1 Message Date
liyizhouAI 67f33a2336 perf: 4x speedup for graph building 
- Increase chunk size 500→2000 (reduces chunks from ~125 to ~32 for 60K docs)
- Use Graphiti add_episode_bulk for parallel processing (5 episodes per batch)
- Fallback to sequential if bulk fails
- Target: 60K doc graph build in ~10-15 min instead of 45 min

Generated with [Claude Code](https://claude.ai/code)
via [Happy](https://happy.engineering)

Co-Authored-By: Claude <noreply@anthropic.com>
Co-Authored-By: Happy <yesreply@happy.engineering>
 2026-04-14 14:09:41 +08:00
liyizhouAI 339f2f2e50 fix: Neo4j DateTime serialization + match all edge types 
- Convert Neo4j DateTime objects to strings via _safe_str() helper
- Query all edge types (RELATES_TO + MENTIONS), not just RELATES_TO
- Fix get_node_edges to match any relationship type

Generated with [Claude Code](https://claude.ai/code)
via [Happy](https://happy.engineering)

Co-Authored-By: Claude <noreply@anthropic.com>
Co-Authored-By: Happy <yesreply@happy.engineering>
 2026-04-14 13:38:11 +08:00
liyizhouAI d5a908077a perf: upgrade Graphiti LLM to Qwen 32B + improve progress feedback 
- Upgrade from Qwen2.5-7B to Qwen2.5-32B-Instruct (faster, better quality)
- Add real-time progress messages: "正在处理第 X/Y 个文本块"
- Log per-episode processing time
- Expand progress range from 15-55% to 15-90% (no more episode polling step)
- Add GRAPHITI_LLM_* separate config for graph building LLM

Generated with [Claude Code](https://claude.ai/code)
via [Happy](https://happy.engineering)

Co-Authored-By: Claude <noreply@anthropic.com>
Co-Authored-By: Happy <yesreply@happy.engineering>
 2026-04-13 15:31:16 +08:00
liyizhouAI a35ba347f3 fix: use SiliconFlow free LLM+embedding for Graphiti 
- Graphiti LLM: SiliconFlow Qwen/Qwen2.5-7B-Instruct (free, supports structured output)
- Graphiti Embedding: SiliconFlow BAAI/bge-m3 (free, OpenAI-compatible)
- MiniMax Coding Plan doesn't support structured output needed by Graphiti
- Separate GRAPHITI_LLM_* config from main LLM_* config
- Remove _wait_for_episodes call (Graphiti processes synchronously)

Generated with [Claude Code](https://claude.ai/code)
via [Happy](https://happy.engineering)

Co-Authored-By: Claude <noreply@anthropic.com>
Co-Authored-By: Happy <yesreply@happy.engineering>
 2026-04-13 14:09:26 +08:00
liyizhouAI 05a5ba0775 fix: add MiniMax embedder and reranker for Graphiti 
- MiniMaxEmbedder: custom EmbedderClient for MiniMax's non-OpenAI-compatible
  embedding API (uses 'texts' field instead of 'input')
- Pass LLM config to OpenAIRerankerClient to avoid OPENAI_API_KEY requirement
- Both embedder and reranker now use MiniMax API credentials

Generated with [Claude Code](https://claude.ai/code)
via [Happy](https://happy.engineering)

Co-Authored-By: Claude <noreply@anthropic.com>
Co-Authored-By: Happy <yesreply@happy.engineering>
 2026-04-13 13:12:30 +08:00
liyizhouAI fff7edce2a feat: migrate from Zep Cloud to self-hosted Graphiti + Neo4j 
Replace Zep Cloud ($25/mo) with open-source Graphiti + Neo4j:
- New graphiti_client.py: unified wrapper with sync bridge for async Graphiti
- Modified graph_builder.py: use Graphiti add_episode instead of Zep batch API
- Modified zep_entity_reader.py: Neo4j Cypher queries replace Zep pagination
- Modified zep_tools.py: Graphiti search replaces Zep Cloud search
- Modified zep_graph_memory_updater.py: Graphiti add_episode replaces Zep add
- Modified oasis_profile_generator.py: GraphitiClient replaces Zep client
- Updated config.py: NEO4J_URI/USER/PASSWORD replace ZEP_API_KEY
- Updated requirements.txt: graphiti-core + neo4j replace zep-cloud
- Added PRD.md: product requirements document with token analysis

Generated with [Claude Code](https://claude.ai/code)
via [Happy](https://happy.engineering)

Co-Authored-By: Claude <noreply@anthropic.com>
Co-Authored-By: Happy <yesreply@happy.engineering>
 2026-04-13 12:07:32 +08:00