n8n Guardrails and Patterns

Use this page when your RAG or agent workflow runs in n8n. It maps common automation failures to the exact structural fixes in the Problem Map and gives a minimal recipe you can paste into a workflow.

Core acceptance

ΔS(question, retrieved) ≤ 0.45
coverage ≥ 0.70 for the target section
λ stays convergent across 3 paraphrases

Typical breakpoints and the right fix

Nodes fire before dependencies are ready
Fix No.14: Bootstrap Ordering → Open
First call after deploy crashes or uses wrong env/secret
Fix No.16: Pre-Deploy Collapse → Open
Circular waits between index build and retriever, or Merge nodes loop forever
Fix No.15: Deployment Deadlock → Open
High vector similarity but wrong meaning
Fix No.5: Embedding ≠ Semantic → Open
Wrong snippet chosen or citations do not line up
Fix No.8: Retrieval Traceability → Open
Contract the payload: Data Contracts → Open
Hybrid retrieval performs worse than a single retriever
Pattern: Query Parsing Split → Open
Also review: Rerankers → Open
Some facts never surface even though indexed
Pattern: Vectorstore Fragmentation → Open
Two sources get merged in the answer
Pattern: Symbolic Constraint Unlock (SCU) → Open

Minimal setup checklist for any n8n flow

Warm-up fence before RAG or LLM nodes
Validate VECTOR_READY == true, INDEX_HASH matches, and secrets exist.
If not ready, short-circuit to a Wait node then retry with a capped counter.
Spec: Bootstrap Ordering
Idempotency and dedupe
Compute dedupe_key = sha256(source_id + revision + index_hash).
Check or write the key using Redis, Postgres, or n8n’s Data Store. Drop duplicates.
RAG boundary contract
Require fields: snippet_id, section_id, source_url, offsets, tokens.
Enforce cite then explain. Forbid cross section reuse.
Specs: Data Contracts · Retrieval Traceability
Observability probes
Log ΔS(question, retrieved). Log λ per step: retrieve, assemble, reason.
Alert when ΔS ≥ 0.60 or λ flips divergent.
Overview: RAG Architecture & Recovery
Concurrency guard
Use a single writer for index updates. Set queue mode or global mutex for write steps.
See: Deployment Deadlock
Regression gate
Require coverage ≥ 0.70 and ΔS ≤ 0.45 before publishing.
Eval: RAG Precision/Recall

n8n recipe you can copy

Replace the concrete nodes with your stack. Keep the guardrails.

Trigger
Fixed source_id and revision. Record wf_rev.
Warm-up Check (Code node)
Pull INDEX_HASH, VECTOR_READY, and secrets.
If not ready, set ready=false.
Branch: Not ready
Wait 30–90 seconds. Increment a retry counter. Stop after N attempts.
Branch: Ready
Retrieval node
- Call retriever with explicit metric and same analyzer as the writer.
- Emit snippet_id, section_id, source_url, offsets, tokens.
  ΔS probe node
- Compute ΔS(question, retrieved). If ΔS ≥ 0.60 set needs_fix=true.
  LLM node
- Model reads TXT OS and follows the WFGY schema. Enforce cite then explain.
  Trace sink
- Store question, snippet_id, ΔS, λ_state, INDEX_HASH, dedupe_key.
  Idempotency guard
- Before side effects, check the KV for dedupe_key. Skip if it already exists.

Copy-paste prompt for the LLM node


I uploaded TXT OS and the WFGY Problem Map files.
This n8n flow retrieved {k} snippets with fields {snippet\_id, section\_id, source\_url, offsets}.
Question: "{user\_question}"

Do:

1. Validate cite-then-explain. If citations are missing, fail fast and return the fix tip.
2. If ΔS(question, retrieved) ≥ 0.60, propose the minimal structural fix referencing:
   retrieval-playbook, retrieval-traceability, data-contracts, rerankers.
3. Return a JSON plan:
   { "citations": \[...], "answer": "...", "λ\_state": "→|←|<>|×", "ΔS": 0.xx, "next\_fix": "..." }
   Keep it auditable and short.

Common n8n gotchas

Set or Move nodes rename fields and break your data contract
Lock field names and run a schema check before the LLM node.
Merge or Split In Batches cause duplicate writes
Add a single writer stage with idempotency keys and a queue.
Cron triggers overlap with long runs
Use a global mutex or skip if dedupe_key already exists.
HyDE prompt built inside the flow differs from the API client
Keep tokenizer and casing identical, or switch to reranking.
See: Rerankers

When to escalate

ΔS stays ≥ 0.60 after chunk and retrieval fixes
Rebuild index with explicit metric and normalization.
See: Retrieval Playbook
Answers alternate across runs with identical input
Investigate memory desync and version skew.
See: Pre-Deploy Collapse

🔗 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	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 —
⭐ WFGY Engine 2.0 is already unlocked. ⭐ Star the repo to help others discover it and unlock more on the Unlock Board.

10 KiB Raw Blame History Unescape Escape