WFGY/ProblemMap/GlobalFixMap/Embeddings/hybrid_retriever_weights.md

Hybrid Retriever Weights — Guardrails and Fix Patterns

Use this page when BM25 plus dense retrieval performs worse than either one alone, when top k flips between runs, or when hybrid recall feels random after an index rebuild. The goal is to put both retrievers on a common score space, then set stable weights, add a single deterministic reranker, and verify with ΔS, coverage, and λ.

Open these first

• Visual map and recovery: rag-architecture-and-recovery.md
• End to end retrieval knobs: retrieval-playbook.md
• Query split root cause: patterns/pattern_query_parsing_split.md
• Reranking for order control: rerankers.md
• Why this snippet and cite first: retrieval-traceability.md
• Embedding vs meaning failures: embedding-vs-semantic.md
• Vector store health: vectorstore_fragmentation.md
• Normalization pitfalls: normalization_and_scaling.md

When to use this page

• BM25 finds the right doc yet dense misses, or the reverse
• Hybrid returns unstable order while single retrievers look stable
• Raising k helps only one side while the other side adds noise
• A client upgrade changes analyzers or tokenization and hybrid breaks

Acceptance targets

• ΔS(question, retrieved) ≤ 0.45
• Coverage of the target section ≥ 0.70
• λ remains convergent across three paraphrases and two seeds
• E_resonance stays flat on long windows

---

Core idea: put scores on the same ruler

Let s_dense be the dense similarity and s_sparse be the BM25 or keyword score. Raw values live in different spaces. Calibrate to a common space, then blend.

Common choices:

• Z score per retriever z = (s − μ) / σ computed on the candidate pool for the current query

• Min max per retriever m = (s − min) / (max − min) on the candidate pool

• Platt style logistic calibration on a gold set Fit p = sigmoid(a s + b) for each retriever

After calibration, blend with a single weight:

S = α * z_dense + (1 − α) * z_sparse Pick α by grid search on a gold set, then lock it.

Always add a single deterministic reranker on the union of candidates. See rerankers.md.

---

60 second fix checklist

1. Normalize both score streams per query. Prefer z score on the union pool.
2. Build a union candidate set. Choose k_dense and k_sparse so that the union size before rerank is about 20 to 50.
3. Grid search α in steps of 0.1 on your gold set. Pick α that maximizes coverage at k and stabilizes λ across seeds.
4. Add a single cross encoder reranker on top of the blended scores. Fix the seed.
5. Lock the header order and cite first to clamp λ while you tune. See retrieval-traceability.md.
6. Verify the targets on three paraphrases and two seeds.

---

Symptom to likely cause

• Hybrid worse than BM25 alone Likely cause. Dense scores unnormalized or wrong metric. Open normalization_and_scaling.md.

• Hybrid worse than dense alone Likely cause. Analyzer mismatch or stopword removal on one path. Open pattern_query_parsing_split.md.

• Order flips between runs Likely cause. Header reorder or union set not stable. Fix header order and use a stable tie break by doc_id then section_id. Open retrieval-traceability.md.

• Right doc appears only when k is very large Likely cause. Fragmented stores or per tenant splits without union. Open vectorstore_fragmentation.md.

---

Minimal calibration recipe

Gold set Ten to fifty questions with known anchor sections. Keep per domain if you support many domains.

Union recall Collect k_dense and k_sparse candidates. Form the union. Store raw scores and features.

Per query normalization Compute z scores on the union for each retriever. Keep the statistics for audits.

Blend and rerank Score with S = α * z_dense + (1 − α) * z_sparse. Apply a single cross encoder rerank with a fixed seed. Resolve ties by doc_id then section_id.

Selection of α Grid search α from 0.0 to 1.0. Choose α that maximizes coverage at k and keeps ΔS ≤ 0.45 on the anchor.

Freeze Record α, k values, normalization method, reranker id, and seed in your contract.

---

Copy paste probes

Probe A — alpha sweep
For α in {0.0..1.0 step 0.1}:
  compute coverage@k and median ΔS on the gold set
Pick α with highest coverage. Break ties by lower ΔS and lower variance.

Probe B — normalization sanity
Toggle {z, minmax, logistic} on the same union pool.
If results change wildly, per query statistics are unstable or candidate pools differ.

Probe C — union size
Sweep k_dense and k_sparse so union size ∈ [20, 50].
If coverage improves up to a point then drops, prune duplicates and near duplicates before rerank.

Probe D — header clamp
Swap harmless header lines. If top k flips, clamp header order and apply cite first. Re test α after clamp.

---

Contract fields to add

{
  "retrievers": {
    "dense": {
      "model": "model-id",
      "metric": "cosine",
      "normalize": "zscore",
      "k": 40
    },
    "sparse": {
      "analyzer_rev": "bm25-v3",
      "normalize": "zscore",
      "k": 200
    }
  },
  "hybrid": {
    "alpha": 0.6,
    "blend": "alpha_sum",
    "tie_break": ["doc_id", "section_id"]
  },
  "reranker": {
    "name": "cross-encoder-id",
    "seed": 7,
    "top_k": 20
  }
}

---

Verification checklist

• Coverage ≥ 0.70 on the gold anchors
• ΔS(question, retrieved) ≤ 0.45
• λ convergent across two seeds and three paraphrases
• Top k overlap across seeds ≥ 0.8 after rerank
• Stable results after harmless header reorders

---

Copy paste prompt for the LLM step

TXT OS and WFGY Problem Map are loaded.

My issue: hybrid BM25 + dense is worse than single retrievers.
Traces:
- zscore_dense=..., zscore_sparse=..., alpha=...
- union_size=..., reranker=name, seed=...
- ΔS(question,retrieved)=..., coverage=..., λ across 3 paraphrases

Tell me:
1) the failing layer and why,
2) the exact WFGY page to open next,
3) a minimal calibration plan for hybrid weights and normalization,
4) a verification plan to reach coverage ≥ 0.70 and ΔS ≤ 0.45.
Use BBMC, BBCR, BBPF, 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” to boot

---

🧭 Explore More

Module	Description	Link
WFGY Core	WFGY 2.0 engine. full symbolic reasoning and math stack	View →
Problem Map 1.0	Initial 16 mode diagnostic and symbolic fixes	View →
Problem Map 2.0	RAG focused failure tree and pipelines	View →
Semantic Clinic Index	Prompt injection. memory bugs. logic drift	View →
Semantic Blueprint	Layer based symbolic reasoning. semantic modulations	View →
Benchmark vs GPT-5	Stress test with full WFGY suite	View →
Starter Village	A guided first run	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 live. ⭐ Star the repo to help others discover it and unlock more on the Unlock Board.