WFGY/ProblemMap/GlobalFixMap/OpsDeploy/vector_index_build_and_swap.md

Vector Index — Build, Validate, and Atomic Swap

🧭 Quick Return to Map

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

• OpsDeploy — operations automation and deployment pipelines
• 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.

A zero-downtime pattern to rebuild your vector index offline, prove it’s correct, then cut traffic over with a single reversible pointer. Works with FAISS, Milvus, Weaviate, Qdrant, pgvector, Redis, Chroma, Vespa, Typesense, and more.

---

Open these first

• RAG map and recovery: rag-architecture-and-recovery.md
• Retrieval knobs end-to-end: retrieval-playbook.md
• Why this snippet (schema): retrieval-traceability.md, data-contracts.md
• Embedding vs meaning: embedding-vs-semantic.md
• Fragmentation failure mode: pattern_vectorstore_fragmentation.md
• Metric mismatch & normalization: metric_mismatch.md, normalization_and_scaling.md

---

When to use

• You change embedding model, dimension, normalization, or distance metric.
• You modify chunk schema, overlaps, or anchor fields.
• You add a reranker or change k, cutoff, weights.
• You suspect index skew or store fragmentation.
• You need a region-by-region cutover with instant rollback.

---

Acceptance targets

• ΔS(question, retrieved) ≤ 0.45 on three paraphrases (gold set).
• Coverage to the correct section ≥ 0.70.
• λ remains convergent across two seeds.
• No schema drift in required fields: {snippet_id, section_id, source_url, offsets, tokens}.
• p95 latency change within ±15% of baseline at k used in prod.

---

Build pipeline (offline)

1. Freeze specs
   Pin EMBED_MODEL_VER, EMBED_DIM, NORM, metric, TOK_VER, ANALYZER_CONF, CHUNK_SCHEMA_VER.
   Record as a manifest next to the index.

2. Re-chunk & re-embed
   Apply your chunking checklist; write vectors with doc ids and anchor metadata.

3. Construct index docs_vB
   Use store-appropriate build parameters. Keep docs_vA live.

4. Attach reranker (if any)
   Persist RERANK_CONF and deterministic ordering parameters.

5. Write integrity probes
   Store INDEX_HASH = hash(all vectors + manifest). Emit alongside the retriever.

---

Validation (before any traffic)

• Gold set eval (20–40 Qs)
  Run baseline docs_vA vs candidate docs_vB. Log ΔS, coverage, λ, latency.

• Anchor triangulation
  Compare ΔS(retrieved, anchor) vs ΔS(retrieved, decoy). If close, fix chunking.

• Metric sanity
  If cosine neighbors look right but meaning is off, re-check metric/norm rules.

• Fragmentation scan
  If top-k distributions differ wildly across partitions, de-frag or rebuild.

• Contracts
  Verify snippet fields and cite-then-explain are intact.

---

Atomic swap patterns

Alias swap (preferred)

Keep a stable read name (e.g., docs_live) and flip alias from docs_vA to docs_vB.
Rollback = flip back to docs_vA.

Config pointer (KV)

Keep a single INDEX_PTR=docs_vX in a one-writer KV. All readers dereference at request start.
Rollback = set back to previous pointer.

Per-region staged swap

Swap one region at a time; watch ΔS/coverage/λ for 15–30 minutes before next region.

---

60-second cutover checklist

1. Candidate docs_vB passes gold eval targets.
2. INDEX_HASH, EMBED_SCHEMA, RERANK_CONF emitted by retriever match the manifest.
3. Caches are primed or partitioned by INDEX_HASH.
4. Canary at 5% is green.
5. Alias flip or pointer update is one operation and reversible.

---

Stop & rollback

• Hard stop if ΔS p95 drift > 0.15 or coverage < 0.60 or λ flip rate > 0.20.
• Hard stop if tool loops or 5xx > 1%.
• Rollback = alias flip back to docs_vA (or pointer revert).
• After rollback, open: debug_playbook.md.

---

Store-specific notes

• FAISS / Chroma: ensure build-time metric matches query-time; re-normalize if switching from dot to cosine.
• Qdrant / Weaviate / Milvus: pin HNSW/IVF params; rebuild rather than mutate when dimension changes.
• pgvector: match vector_l2_ops/vector_cosine_ops with embedding norm; verify ANALYZER_CONF if paired with text search.
• Redis: keep alias via FT.ALIASADD; avoid multi-writer FT.CREATE races.
• Vespa / Typesense: schema and field types versioned; perform shadow feed before activation.

---

Pseudo commands

Build

# generate chunks and vectors
wfgy_chunk --schema s128/o32 corpus.jsonl > chunks.jsonl
wfgy_embed --model text-embedding-3-large chunks.jsonl > vectors.fbin

# build candidate index
wfgy_index build --metric cosine --norm l2 --in vectors.fbin --out docs_vB

# compute and store manifest + hash
wfgy_index manifest --index docs_vB > manifest.json
wfgy_index hash --index docs_vB > INDEX_HASH.txt

Validate vs baseline

wfgy_eval rag --gold gold_40.json \
  --indexes docs_vA,docs_vB \
  --targets "ds<=0.45,cov>=0.70,lambda=convergent"

Swap and rollback (alias style)

vec alias update docs_live --to docs_vB
# rollback
vec alias update docs_live --to docs_vA

---

Common pitfalls

• Mixed analyzers/tokenizers across ingest vs query.
• “Minor” embedding model update that changes dimension or norm assumptions.
• Reranker cutoff mismatch between staging and prod.
• Cache keys without INDEX_HASH causing stale blends.
• Two writers touching the same live alias.

---

🔗 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 — 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.