Local extract reliability (Ollama / LM Studio)

Local extract reliability (Ollama / LM Studio)

Quick ops runbook when KG extraction stalls with Network error … /api/chat or Local inference gate saturated under a single-slot Ollama runner (-np 1).

Recommended local profile

Knob	Value	Why
OLLAMA_CONTEXT_LENGTH	8192	Avoid 128k runner cost on extract
EDGEQUAKE_MAX_CONCURRENT_EXTRACTIONS	1	Match Ollama serial slot
EDGEQUAKE_PROVIDER_BUDGET / EDGEQUAKE_LOCAL_MAX_INFLIGHT	1	Gate admits one in-flight chat
EDGEQUAKE_EXTRACT_REASONING_EFFORT	none	Forces think:false / reasoning:"off"
Extract model	gemma4:latest (or cloud)	Prefer over 35B for bulk PDFs

make dev / make backend-bg export these defaults for ollama/lmstudio providers.

Unblock a stuck document

1. Restart backend so think-off + admission clamps are live:

   make stop
   export OLLAMA_CONTEXT_LENGTH=8192
   export EDGEQUAKE_MAX_CONCURRENT_EXTRACTIONS=1
   export EDGEQUAKE_PROVIDER_BUDGET=1
   export EDGEQUAKE_EXTRACT_REASONING_EFFORT=none
   make backend-bg   # or make dev

2. Prefer a smaller extract model on the workspace (Settings → models) if still on qwen3.6:35b*.
3. Cancel the stuck track in Documents → Active run, then requeue / re-upload.
4. Confirm health: curl -s http://localhost:8090/health | jq .providers

Success signals

• Near-zero Network error under single-doc local extract with healthy Ollama.
• Gate saturation becomes wait + heartbeat (Extracting … in flight), not connection storms.
• Outbound Ollama body includes "think": false and options.num_ctx ≤ configured ctx.