WFGY/ProblemMap/GlobalFixMap/VectorDBs_and_Stores/elasticsearch.md

Elasticsearch: Guardrails and Fix Patterns

🧭 Quick Return to Map

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

• VectorDBs_and_Stores — vector indexes and storage backends
• 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 compact field guide to stabilize Elasticsearch vector search when your RAG or agent stack loses accuracy. Use this to localize the failing layer and jump to the exact WFGY fix page.

Open these first

• Visual map and recovery: RAG Architecture & Recovery
• End-to-end retrieval knobs: Retrieval Playbook
• Why this snippet and trace schema: Retrieval Traceability
• Ordering after recall: Rerankers
• Embedding versus meaning: Embedding ≠ Semantic
• Chunk boundaries and illusions: Hallucination
• Long chains and drift checks: Context Drift, Entropy Collapse
• Structural collapse and recovery: Logic Collapse
• Snippet and citation schema: Data Contracts
• Vector metrics pitfalls: Vectorstore Metrics & FAISS Pitfalls
• Fragmented stores: Vectorstore Fragmentation
• Hybrid query split: Query Parsing Split
• Ops live checks: Live Monitoring for RAG, Debug Playbook

Fix in 60 seconds

1. Measure ΔS
   Compute ΔS(question, retrieved) and ΔS(retrieved, expected anchor).
   Thresholds: stable < 0.40, transitional 0.40–0.60, risk ≥ 0.60.

2. Probe with λ_observe
   Sweep k in {5, 10, 20}. If ΔS is flat high, suspect metric, mapping, or index mismatch.
   Reorder prompt headers. If ΔS spikes, lock schema with Data Contracts.

3. Apply the module
   Retrieval drift → BBMC + Data Contracts.
   Reasoning collapse → BBCR bridge + BBAM variance clamp.
   Dead ends in long runs → BBPF alternate path.

4. Verify acceptance
   Coverage ≥ 0.70 to target section.
   ΔS ≤ 0.45 on three paraphrases.
   λ remains convergent. Traces reproducible.

Elasticsearch breakpoints and the right repair

1) dense_vector mapping mismatch

Symptoms: insert errors, silent truncation, or chaotic top-k for new docs only.
Why: vector dims do not match encoder, or field not indexed for KNN.
Fix: set type: dense_vector, correct dims, and index: true for KNN. Re-index changed spans. See Data Contracts.

2) Distance metric and normalization

Symptoms: high similarity yet wrong meaning; ordering flips across runs.
Why: using similarity: l2_norm with cosine-trained embeddings, or dot_product without unit-norm vectors.
Fix: align metric to the encoder; normalize for cosine/dot as policy. If you switch, rebuild the index. See Embedding ≠ Semantic and Vectorstore Metrics & FAISS Pitfalls.

3) HNSW underfit and candidate window

Symptoms: gold chunk appears only at very large k.
Why: small m or ef_construction, and num_candidates too low at query time.
Fix: tune m, raise ef_construction, then sweep num_candidates at query. Validate with a reranker. See Retrieval Playbook and Rerankers.

4) knn vs script_score confusion

Symptoms: inconsistent scores between approximate knn and exact script scoring; hybrids regress.
Fix: use exact script_score on a canary set to bound max recall, then tune knn to approach it. Keep one scoring policy in production. Map to Retrieval Traceability.

5) Filters with KNN

Symptoms: empty results when adding filters; massive latency spikes.
Why: pre-filter not supported in your version or filter path not indexed.
Fix: ensure filtered fields are indexed and typed, test post-filter rerank, and document the path in the contract. See Data Contracts.

6) Analyzer drift for hybrid BM25 + vector

Symptoms: hybrid performs worse than either branch alone.
Why: default analyzers, stopwords, or stemming distort lexical scores.
Fix: do not just reuse defaults. Lock analyzers per field and choose explicitly — e.g., icu_tokenizer for Unicode, edge_ngram for prefix search, asciifolding for normalization. Normalize hybrid weights and fuse post-retrieval with a reranker. See Query Parsing Split and Rerankers.

7) Shards, replicas, and refresh

Symptoms: fresh writes never appear; nodes return different sets.
Fix: confirm refresh policy, replica sync, and routing. Add a semantic boot fence before first prod call. See Bootstrap Ordering and Pre-Deploy Collapse.

8) Alias routing and multi-index fragmentation

Symptoms: global recall ok but weak per-alias top-k.
Why: many tiny indices split the neighborhood; wrong read/write alias.
Fix: consolidate to an authoritative index with a facet, fix aliases, rebuild, then rerank. See Vectorstore Fragmentation.

9) Upsert hygiene

Symptoms: duplicates, stale docs, toggling answers.
Fix: deterministic IDs, doc_sha in metadata, idempotent loaders, periodic dedupe. Validate with golden queries. See Retrieval Traceability.

Observability probes

• k-sweep curve: run k in 5, 10, 20. Flat high ΔS → metric, mapping, or index fault.
• Exact vs approx: compare script_score exact against knn. Large gap → retune HNSW and num_candidates.
• Hybrid toggle: vector only vs hybrid. If hybrid regresses, repair analyzers and fusion weights.
• Reranker audit: strong reranker should reduce ΔS while recall rises. If not, rebuild.

Escalate when

• ΔS stays ≥ 0.60 on golden questions after metric, mapping, and HNSW fixes.
• Coverage cannot reach 0.70 even with reranker and anchors.
• Writes appear in logs but remain invisible across shards or replicas.

Open:

• Retrieval Playbook
• Data Contracts
• Live Monitoring for RAG

Copy-paste prompt for your AI

I uploaded TXT OS and the WFGY Problem Map files.

Target system: Elasticsearch.

* symptom: \[brief]
* traces: ΔS(question,retrieved)=..., ΔS(retrieved,anchor)=..., λ states
* mapping: \[field, dims, index=true, similarity=cosine|dot\_product|l2\_norm]
* knn: \[k, num\_candidates, hnsw m, ef\_construction]
* exact: \[script\_score policy if used]
* hybrid: \[match/bm25 fields, analyzers, weights]
* ingest: \[ids, doc\_sha, upsert policy]
* routing: \[index/alias, shards, replicas, refresh]

Tell me:

1. which layer is failing and why,
2. which exact fix page to open from this repo,
3. minimal steps to push ΔS ≤ 0.45 and keep λ convergent,
4. how to verify with a reproducible test.

Use BBMC/BBPF/BBCR/BBAM when relevant.

---

🔗 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 and semantic modulations	View →
Benchmark vs GPT-5	Stress test with full WFGY reasoning suite	View →
Starter Village	New here. Start with a guided tour	Start →

---

👑 Early Stargazers: See the Hall of Fame
⭐ WFGY Engine 2.0 is live. ⭐ Star the repo to help others discover it and unlock more on the Unlock Board.