| # MEMORY.md — GraphRAG Agent Knowledge Base |
|
|
| ## Learned Facts |
|
|
| ### GraphRAG Performance Characteristics |
| - GraphRAG achieves +21% F1 improvement over baseline RAG on bridge-type questions (HotpotQA) |
| - GraphRAG uses ~2.5x more tokens per query than baseline RAG |
| - Adaptive routing eliminates token overhead for simple queries (complexity < 0.6) |
| - Schema-bounded extraction reduces entity extraction cost by ~90% vs unconstrained |
| - Multi-hop traversal (2 hops) is the sweet spot — 3+ hops adds noise without proportional accuracy gain |
|
|
| ### Provider Performance Notes |
| - Claude Sonnet 4: Best at structured entity extraction (JSON mode via tool_use) |
| - GPT-4o-mini: Best cost/quality ratio for answer generation |
| - Gemini 2.0 Flash: Fastest response times, good for keyword extraction |
| - Ollama llama3.2: Acceptable quality for entity extraction, zero cost |
| - Groq llama-3.3-70b: Near-cloud quality at very low latency (LPU hardware) |
| - DeepSeek R1: Excellent reasoning quality but slower |
| |
| ### TigerGraph Best Practices |
| - Batch upsert limit: 10,000 vertices per call on free tier |
| - GSQL query compilation: 30-120 seconds (install once, run many times) |
| - Vector search is brute-force cosine similarity on free tier (no HNSW index) |
| - Entity deduplication via hash(name.lower() + type.lower()) is essential |
| |
| ### HotpotQA Dataset Notes |
| - Bridge questions: require connecting information across 2 documents |
| - Comparison questions: require comparing attributes of 2 entities |
| - Supporting facts: gold standard for context evaluation |
| - Distractor setting: 8 distractor passages + 2 relevant passages per question |
| |
| ## User Preferences |
| - Prefer concise answers with explicit evidence |
| - Show graph reasoning paths for complex queries |
| - Always display token counts and costs |
| - Default to adaptive routing enabled |
| |
