FAISS: Guardrails and Fix Patterns

A compact repair guide for FAISS retrieval stacks. Use this when recall looks fine but meaning drifts, or when IVF/HNSW tuning flips answers across seeds. The checks below route you to the exact WFGY fix pages and give a minimal recipe you can paste into a runbook.

Open these first

Visual map and recovery: RAG Architecture & Recovery
End-to-end retrieval knobs: Retrieval Playbook
Traceability schema: Retrieval Traceability
Embedding ≠ meaning: Embedding vs Semantic
Vectorstore health: Vectorstore Fragmentation
Query split (HyDE vs BM25): Query Parsing Split
Ordering control: Rerankers
Snippet and citation schema: Data Contracts
Evaluating RAG: RAG Precision/Recall
FAISS specifics: Vectorstore Metrics & FAISS Pitfalls
Live checks: Live Monitoring

Fix in 60 seconds

Measure ΔS
- Compute ΔS(question, retrieved) and ΔS(retrieved, expected anchor).
- Thresholds: stable < 0.40, transitional 0.40–0.60, risk ≥ 0.60.
Probe with λ_observe
- Sweep k ∈ {5, 10, 20} and for IVF sweep nprobe ∈ {1, 4, 8, 16}.
- For HNSW, sweep efSearch ∈ {32, 64, 128}.
- If ΔS flattens high across k, suspect metric/index mismatch.
Apply the module
- Retrieval drift → BBMC + Data Contracts.
- Reasoning collapse after retrieval → BBCR bridge + BBAM variance clamp, see Logic Collapse.
- Dead ends in long runs → BBPF alternate path + Context Drift.
Verify
- Coverage to target section ≥ 0.70, ΔS ≤ 0.45 on three paraphrases, λ stays convergent across seeds.

Typical breakpoints and the right fix

Symptom	Likely cause	Open this	Minimal fix
High cosine similarity but wrong meaning	IP vs L2 mixup, un-normalized embeddings	Embedding vs Semantic	Normalize vectors; match metric to embedder; re-index
Good recall, messy top-k order	Rerank missing or weak	Rerankers	Add cross-encoder rerank, k=50→top-10
Some facts never show up	Shards or label fragmentation	Vectorstore Fragmentation	Merge shards; rebuild IVF lists; verify dim
Answers flip between runs	IVF `nlist/nprobe` underfit, PQ over-aggressive	FAISS Pitfalls	Raise `nprobe`, enlarge training set, reduce PQ
Hybrid gets worse than single retriever	Query split and prompt coupling	Query Parsing Split	Split semantic vs lexical prompts; fuse post-retrieval

FAISS quick checklist

Confirm dimension matches the embedding model output exactly.
Confirm metric: IP with normalized vectors, or L2 with raw vectors. Do not mix.
For IVF, set nlist based on corpus size, train with at least 100× nlist examples.
Start with nprobe ≈ sqrt(nlist) and tune upward until ΔS stabilizes.
For HNSW, raise efConstruction and efSearch until ΔS stops improving.
Rebuild the index after changing normalization or metric.
Lock the snippet schema and citations using Data Contracts.

Copy-paste repair prompt


audit FAISS retrieval with ΔS and λ\_observe.
report: metric choice (IP/L2), normalization, dim, index type, nlist/nprobe or HNSW ef.
run three paraphrases, k in {5,10,20}. if ΔS stays >0.45, switch to normalized IP and rebuild.
apply BBMC + Data Contracts; add reranker for top-50→10. show before/after ΔS table.

Acceptance targets

Coverage ≥ 0.70 to the target section.
ΔS ≤ 0.45 across three paraphrases.
λ remains convergent across seeds.
E_resonance flat under long windows.

🔗 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 + ”
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.

8.8 KiB Raw Blame History Unescape Escape