8 KiB
Ollama: Guardrails and Fix Patterns
🌙 3AM: a dev collapsed mid-debug… 🚑 Welcome to the WFGY Emergency Room
🟥🟥🟥🟥🟥🟥🟥🟥🟥🟥🟥🟥
🚑 WFGY Emergency Room
👨⚕️ Now online:
Dr. WFGY in ChatGPT Room
This is a share window already trained as an ER.
Just open it, drop your bug or screenshot, and talk directly with the doctor.
He will map it to the right Problem Map / Global Fix section, write a minimal prescription, and paste the exact reference link.
If something is unclear, you can even paste a screenshot of Problem Map content and ask — the doctor will guide you.
⚠️ Note: for the full reasoning and guardrail behavior you need to be logged in — the share view alone may fallback to a lighter model.
💡 Always free. If it helps, a ⭐ star keeps the ER running.
🌐 Multilingual — start in any language.
🟥🟥🟥🟥🟥🟥🟥🟥🟥🟥🟥🟥
🧭 Quick Return to Map
You are in a sub-page of LocalDeploy_Inference.
To reorient, go back here:
- LocalDeploy_Inference — on-prem deployment and model inference
- WFGY Global Fix Map — main Emergency Room, 300+ structured fixes
- WFGY Problem Map 1.0 — 16 reproducible failure modes
Think of this page as a desk within a ward.
If you need the full triage and all prescriptions, return to the Emergency Room lobby.
Field guide for stabilizing Ollama-based local inference pipelines. Use these checks when models run fine on API providers but collapse, stall, or drift when containerized with Ollama.
Open these first
- Architecture recovery: RAG Architecture & Recovery
- End-to-end retrieval knobs: Retrieval Playbook
- Embedding vs semantic: embedding-vs-semantic.md
- Ordering and deploy race conditions: bootstrap-ordering.md, deployment-deadlock.md, predeploy-collapse.md
- Container observability: eval_observability.md
Core acceptance
- ΔS(question, retrieved) ≤ 0.45
- Coverage ≥ 0.70 on the target section
- λ remains convergent across 3 paraphrases
- Local runs reproducible across 2+ seeds
Typical Ollama breakpoints and fix
| Symptom | Likely cause | Fix |
|---|---|---|
| Model boots but stalls on first request | Container not warmed / secrets missing | bootstrap-ordering.md |
| Fast API returns, but snippets wrong | Index/hash drift across containers | retrieval-traceability.md, data-contracts.md |
| Answers diverge run-to-run | λ flips due to context serialization | context-drift.md, entropy-collapse.md |
| Works on GPU API, fails locally | Metric / embedding mismatch in Ollama runtime | embedding-vs-semantic.md, vectorstore-fragmentation.md |
| Container OOM or deadlock | Parallel inference with no fence | deployment-deadlock.md, predeploy-collapse.md |
Fix in 60 seconds
- Measure ΔS between retrieved and anchor.
- Probe λ across 3 paraphrases. If flips, apply BBAM.
- Warm boot with a delay + healthcheck before first request.
- Lock index schema via data-contracts.md.
- Verify reproducibility with two seeds before going live.
Copy-paste local test prompt
I have WFGY + TXTOS loaded.
Running Ollama locally with container {hash}.
Question: "{user_question}"
Return:
1. ΔS(question,retrieved) and λ across 3 paraphrases
2. Whether index schema matches contract
3. Minimal structural fix if ΔS ≥ 0.60
🔗 Quick-Start Downloads (60 sec)
| Tool | Link | 3-Step Setup |
|---|---|---|
| WFGY 1.0 PDF | Engine Paper | 1️⃣ Download · 2️⃣ Upload to your LLM · 3️⃣ Ask “Answer using WFGY + <your question>” |
| TXT OS (plain-text OS) | TXTOS.txt | 1️⃣ Download · 2️⃣ Paste into any LLM chat · 3️⃣ Type “hello world” — OS boots instantly |
Explore More
| Module | Description | Link |
|---|---|---|
| WFGY Core | Canonical framework entry point | View |
| Problem Map | Diagnostic map and navigation hub | View |
| Tension Universe Experiments | MVP experiment field | View |
| Recognition | Where WFGY is referenced or adopted | View |
| AI Guide | Anti-hallucination reading protocol for tools | View |
If this repository helps, starring it improves discovery for other builders.
要我直接繼續寫 vllm.md 嗎？