Graphiti hybrid search misfires on domain-specific memories, even when the LLM gets it right
Opposite_Courage_531 · reddit · 2026-07-26
The author is asking for advice on improving retrieval for domain-specific memory queries using Graphiti and hybrid search.
They built synthetic memories, pushed them into Neo4j, and ran a small A/B-style evaluation across six cases. The results show that the system can do well on exact project constraints and historical exclusions, but hybrid retrieval still gets tripped up by lexical overlap. In one example, a query about a lock-free ring buffer for telemetry retrieved an audio-plugin ring-buffer distractor instead of the semantically correct memory.
The post includes the evaluation table and a code snippet, and the author says the LLM itself parsed the injected memory correctly — the weakness is in retrieval, not generation. They’re looking for ideas to improve domain-specific search logic.
More from coding & agent
- Most agent teams should tune the harness before touching model weights — hwchase17 · 2026-07-26
- Model migrations should prove behavior, not just API compatibility — hwchase17 · 2026-07-26
- Owning AI also means owning the data that powers it — hwchase17 · 2026-07-26
- Jeff Dean says context engineering beats raw model size at YC Startup School — ycombinator · 2026-07-26
- Terminal-Bench Score Reaches 60% Running on 56GB VRAM with Single RTX Pro 6000 — victormustar · 2026-07-26
- A 4GB laptop GPU can run a local Qwen agent stack, but RAG forces model swaps — maikerukonare · 2026-07-26