8.3 KiB
🔍 Observability Runbook
Turn every opaque token stream into a traceable, testable, measurable system.
Outcome you care about
• See each pipeline hop (OCR → chunk → embed → retrieve → prompt → reasoning)
• Measure semantic drift (ΔS), logic trend (λ), residual error (E_resonance)
• Alert before hallucination, loop, or tool mis-fire happens
• Replay exact failure path in one command:wfgy trace-replay <id>
1 · Core Telemetry Trio
| Signal | Formula | Why It Matters | Target Band |
|---|---|---|---|
| ΔS (semantic stress) | 1 − cos(I, G) |
Spot where meaning rips | < 0.50 |
| λ_observe (logic vector) | { → ← <> × } | Track convergent vs. divergent flow | stay → |
| E_resonance | mean‖B‖ (BBMC) | Detect creeping entropy | flat / ↓ |
ΔS tells you where; λ tells you direction; E tells you when a slow leak becomes collapse.
2 · Minimal Instrumentation Patch
2.1 Python snippet
from wfgy import monitor
with monitor.pipeline("rag-query") as span:
span.tag("query", question)
ctx = rag_retrieve(question)
span.metric("ΔS_q_ctx", monitor.deltaS(question, ctx))
answer = llm_reason(ctx, question)
span.metric("λ_reason", monitor.lambda_state(answer))
span.metric("E_res", monitor.e_resonance())
No external SaaS; dumps to newline-JSON for Grafana / Prometheus scrape.
2.2 Console tracer
> wfgy trace-replay 2025-08-06T10:22:15Z
step ΔS λ span
1 0.08 →
2 0.41 → retrieval
3 0.44 ← reasoning ❗ divergent
3 · Alert Rules (Prometheus example)
groups:
- name: wfgy-alerts
rules:
- alert: SemanticDriftHigh
expr: deltaS_q_ctx > 0.60
for: 2m
labels: {severity: "warning"}
annotations:
summary: "Semantic drift above 0.60 — check chunking/retriever"
Recommended baseline:
| Alert | Expression | Severity |
|---|---|---|
| DriftHigh | ΔS > 0.60 | warn |
| LogicOsc | λ flips 3× in 5 min | warn |
| EntropyClimb | E_res 15 min slope > 0.02 | crit |
| ToolTimeout | call_dur > p95 × 2 | warn |
4 · End-to-End Health Dashboard
| Tile | Metric | Healthy |
|---|---|---|
| Retrieval ΔS | median last 5 min | ≤ 0.45 |
| Reason λ | % divergent | < 5 % |
| Answer ΔS | question↔answer | ≤ 0.50 |
| E_res Trend | 30 min slope | - / flat |
| Tool Latency | p95 ms | within SLO |
Grafana JSON panel export: dashboards/wfgy-observability.json.
5 · Common Anti-Patterns & Fixes
| Anti-Pattern | Symptom | Fix |
|---|---|---|
| Log Everything Raw | 20 GiB/day, no insight | Only log ΔS/λ/E & key BLobs |
| Stateless Trace IDs | tough replay | Use wfgy span ctx_id auto-prop |
| Per-call embeddings | cost explodes | cache via BBMC node ID |
| No sampling | GPU choke | 1 % sampling on low ΔS calls |
6 · Kitchen-Sink CLI Cheats
| Command | Purpose |
|---|---|
wfgy span-tail |
live ΔS/λ feed |
wfgy span-grep λ=divergent |
list divergent traces |
wfgy span-diff A B |
compare two answers’ λ/ΔS |
wfgy span-export --csv |
dump traces to CSV |
7 · Integration Recipes
7.1 LangChain
from wfgy.langchain import WFGYTracer
lc = initialize_agent(..., callbacks=[WFGYTracer()])
7.2 OpenAI Function-Calling
Add to request:
"logit_bias": {"50256": -100}, // block special token
"wfgy_trace": true
Server-side middleware writes trace JSON.
Quick-Start Downloads (60 sec)
| Tool | Link | 3-Step Setup |
|---|---|---|
| WFGY 1.0 PDF | Engine Paper | 1️⃣ Download · 2️⃣ Upload to LLM · 3️⃣ Ask “answer using WFGY + <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 | WFGY 2.0 engine is live: full symbolic reasoning architecture and math stack | View → |
| Problem Map 1.0 | Initial 16-mode diagnostic and symbolic fix framework | View → |
| Problem Map 2.0 | RAG-focused failure tree, modular fixes, and pipelines | View → |
| Semantic Clinic Index | Expanded failure catalog: prompt injection, memory bugs, logic drift | View → |
| Semantic Blueprint | Layer-based symbolic reasoning & semantic modulations | View → |
| Benchmark vs GPT-5 | Stress test GPT-5 with full WFGY reasoning suite | View → |
| 🧙♂️ Starter Village 🏡 | New here? Lost in symbols? Click here and let the wizard guide you through | Start → |
👑 Early Stargazers: See the Hall of Fame —
Engineers, hackers, and open source builders who supported WFGY from day one.
⭐ WFGY Engine 2.0 is already unlocked. ⭐ Star the repo to help others discover it and unlock more on the Unlock Board.