WFGY/ProblemMap/GlobalFixMap/Eval/eval_rag_precision_recall.md

Eval RAG Precision & Recall — Guardrails and Fix Patterns

🧭 Quick Return to Map

You are in a sub-page of Eval.
To reorient, go back here:

• Eval — model evaluation and benchmarking
• 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.

Evaluation disclaimer (RAG precision and recall)
Precision and recall here are computed in a controlled RAG scenario with specific data and judgement rules.
They should be used to debug retrieval behavior, not as general claims about model intelligence.

---

This page defines how to measure precision and recall in RAG pipelines under the WFGY framework. It sets acceptance thresholds, common pitfalls, and structural fixes to keep evaluations meaningful and reproducible.

---

Open these first

• Visual map and recovery: RAG Architecture & Recovery
• Retrieval contract: Data Contracts
• Traceability schema: Retrieval Traceability
• Embedding drift: Embedding ≠ Semantic
• Hallucination boundaries: Hallucination

---

Acceptance targets

• Precision ≥ 0.75 at citation level
• Recall ≥ 0.70 of gold anchor snippets
• ΔS(question, retrieved) ≤ 0.45 for majority of pairs
• λ remains convergent across 3 paraphrases and 2 random seeds
• Evaluations must be auditable & reproducible with JSON logs

---

Why precision/recall break in RAG

1. Goldset drift Anchors no longer align with the corpus after updates. → Fix: refresh goldsets with goldset_curation.md.

2. Retrieval contract missing Snippet payloads do not include section IDs or offsets. → Fix: enforce Data Contracts.

3. Precision false positives Semantically near matches but wrong factual anchor. → Fix: rerank with Rerankers.

4. Recall false negatives Correct snippet exists but chunking or index prevents surfacing. → Fix: re-chunk corpus with chunking-checklist.md.

5. Evaluation noise Different seeds or paraphrases give unstable results. → Fix: clamp λ variance with variance_and_drift.md.

---

Quick workflow

1. Load goldset Each gold QA item must cite snippet_id, section_id, source_url.

2. Run retrieval Collect top-k results (k = 5, 10, 20).

3. Log ΔS & λ For each query and paraphrase, record ΔS values and λ states.

4. Compute metrics

   • Precision = correct citations / total citations
   • Recall = correct citations / gold references

5. Regression gate Block deploy if precision < 0.75 or recall < 0.70.

---

Example JSON log

{
  "question": "What causes hallucination re-entry?",
  "gold": ["hallucination-reentry"],
  "retrieved": ["hallucination-reentry", "entropy-drift"],
  "precision": 0.50,
  "recall": 1.00,
  "ΔS": 0.38,
  "λ_state": "→"
}

---

Common pitfalls

• Evaluating only precision → recall collapses unnoticed.
• Counting fuzzy hits as correct → ΔS may be high, but factually wrong.
• No paraphrases tested → λ instability hidden.
• Relying on one seed → fragile numbers that don’t generalize.

---

🔗 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.