Microsoft Power Automate — Guardrails and Fix Patterns

Use this when your workflow is built with Power Automate (cloud flows, AI Builder, custom connectors) and you see wrong citations, unstable answers, mixed sources, or silent failures that “look green” in run history.

Acceptance targets

ΔS(question, retrieved) ≤ 0.45
Coverage ≥ 0.70 to the intended section/record
λ remains convergent across 3 paraphrases

Typical breakpoints → exact fixes

Output is plausible yet cites the wrong doc/snippet
Fix No.1: Hallucination & Chunk Drift →
Hallucination ·
Retrieval Playbook
Vector similarity looks fine, meaning is off
Fix No.5: Embedding ≠ Semantic →
Embedding ≠ Semantic
Some facts exist in SharePoint/Dataverse but never surface in top-k
Pattern: Vectorstore Fragmentation →
Vectorstore Fragmentation
Can’t explain why a snippet was chosen; run history shows only final text
Fix No.8: Retrieval Traceability with snippet schema →
Retrieval Traceability ·
Data Contracts
Long chains or approvals flatten tone and drift logically
Fix No.3/No.9: Context Drift and Entropy Collapse →
Context Drift ·
Entropy Collapse
Flow passes in Test but fails after environment or connection swap
Infra: Pre-Deploy/Bootstrap/Deadlock →
Pre-Deploy Collapse ·
Bootstrap Ordering ·
Deployment Deadlock
Confident tone with wrong answers in AI Builder actions
Fix No.4: Bluffing/Overconfidence →
Bluffing

Minimal Power Automate pattern with WFGY checks

Below is a compact flow outline. It enforces cite-first schema, observable retrieval, and a ΔS/λ post-check.

Trigger: When an HTTP request is received
Actions:
1) Initialize variable "k" = 10
2) Parse JSON "question" from request
3) HTTP → your retriever endpoint
   - Method: POST
   - Body: { "question": "@{variables('question')}", "k": "@{variables('k')}" }
4) Compose "context" = join(retrieved.snippets)
5) Compose "prompt" =
   SYSTEM: Cite lines before any explanation.
   TASK: Answer the user's question using the provided context.
   CONSTRAINTS:
   - Do not mix sources
   - Provide snippet_id for each citation
   CONTEXT:
   @{outputs('Compose_context')}
   QUESTION:
   @{variables('question')}
6) AI Builder / Custom Connector → LLM with "prompt"
7) HTTP → wfgyCheck (custom Azure Function)
   - Body: { "question": "@{variables('question')}",
             "context": "@{outputs('Compose_context')}",
             "answer": "@{outputs('LLM_action')}" }
8) Condition:
   If deltaS ≥ 0.60 OR lambda != "→"
      → Terminate flow (Warn) "High semantic stress. See trace log."
   Else
      → Return 200 with { answer, deltaS, lambda, coverage, citations[] }

What this enforces

Retrieval parameters are explicit and logged in flow run details.
Prompt is schema-locked with cite-first.
WFGY check runs after generation and can fail fast when ΔS is high or λ flips divergent.
Trace table (snippet_id ↔ citation) is returned for audit.

Reference specs RAG Architecture & Recovery · Retrieval Playbook · Retrieval Traceability · Data Contracts

Power Automate specific gotchas

Environment or connection drift: different Dataverse/SharePoint connections between ingestion and query. Pin connections per environment and re-verify secrets. See Pre-Deploy Collapse
Throttling/parallel branches change ordering of records. Add a rerank stage only after per-source ΔS ≤ 0.50. See Rerankers
Parse JSON actions silently drop fields, breaking snippet_id propagation. Validate schemas and keep snippet_id mandatory. See Retrieval Traceability
Embedding metric mismatch between ingestion code (Azure Function/Logic App) and query side. Normalize vectors and pin cosine vs. inner product. See Embedding ≠ Semantic
Scheduled flows rebuild indices unintentionally. Make builds idempotent and gate by boot checks. See Bootstrap Ordering

When to escalate

ΔS remains ≥ 0.60 after chunking and retrieval fixes Work through the playbook, then rebuild the index with explicit metric flags and unit normalization. Retrieval Playbook
Answers flip between Dev/UAT/Prod Verify version skew, connection references, and secrets. 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.

9.7 KiB Raw Blame History Unescape Escape