Store-Agnostic Guardrails for Retrieval

Use this page to harden retrieval quality without changing your vector store. The checks localize failure causes and route you to the exact structural fix so you can verify with measurable targets.

Acceptance targets

ΔS(question, retrieved) ≤ 0.45
Coverage of target section ≥ 0.70
λ remains convergent across 3 paraphrases and 2 seeds
E_resonance stays flat on long windows

15-minute triage checklist

Lock metrics and analyzers
One analyzer for write and read. Verify distance metric and normalization.
Open: Retrieval Playbook
Enforce the snippet contract
Required fields: snippet_id, section_id, source_url, offsets, tokens.
Open: Data Contracts
Trace why this snippet
Add cite-then-explain and store the trace.
Open: Retrieval Traceability
Probe ΔS and λ
Three paraphrases and two seeds. If ΔS ≥ 0.60 or λ flips, clamp variance.
Open: deltaS_probes.md
k sweep and rerankers
k in {5, 10, 20}. Try a deterministic reranker when order matters.
Open: Rerankers · hybrid_reranker_recipe.md
Check chunk boundaries and anchors
If facts exist but never surface, realign chunking and anchors.
Open: chunking-checklist.md · chunk_alignment.md
Detect fragmentation
If coverage is low while index looks healthy, suspect store fragmentation.
Open: pattern_vectorstore_fragmentation.md
Hybrid failure
If hybrid underperforms a single retriever, split parsing and rebalance.
Open: pattern_query_parsing_split.md
Embedding vs meaning
High similarity yet wrong answer means metric or family mismatch.
Open: embedding-vs-semantic.md

Minimal instrumentation you can paste

# Pseudocode: keep these checkpoints store agnostic
def retrieve(q, k=10):
    # unified analyzer and explicit metric
    return retriever.invoke(q, k=k)

def trace_schema(snippet):
    assert {"snippet_id","section_id","source_url","offsets","tokens"} <= set(snippet.keys())

def observe(q, snippets, answer):
    # compute ΔS and λ, record probes
    log = probes.compute(q, snippets, answer)
    if log["ΔS"] >= 0.60 or log["λ_flip"]:
        raise Exception("High ΔS or λ flip. Apply variance clamp and rerankers.")
    return log

def pipeline(q):
    s = retrieve(q, k=10)
    for x in s: trace_schema(x)
    msg = prompt.cite_then_explain(q, s)
    ans = llm.invoke(msg)
    return observe(q, s, ans)

Copy-paste LLM prompt

You have TXT OS and the WFGY pages loaded.

Task:
1) Enforce cite-then-explain with fields {snippet_id, section_id, source_url, offsets, tokens}.
2) Log ΔS(question, retrieved) and λ across 3 paraphrases and 2 seeds.
3) If ΔS ≥ 0.60 or λ flips, propose the smallest structural change referencing:
   retrieval-playbook, retrieval-traceability, data-contracts, rerankers, query-parsing-split.
4) Return JSON:
{ "citations": [...], "answer": "...", "ΔS": 0.xx, "λ_state": "<>", "coverage": 0.xx, "next_fix": "..." }

Symptoms → exact structural fix

Symptom	Likely cause	Open this
High similarity yet wrong meaning	metric or embedding family mismatch	embedding-vs-semantic.md
Facts exist but never retrieved	chunk drift or store fragmentation	chunking-checklist.md · pattern_vectorstore_fragmentation.md
Hybrid worse than single retriever	query parsing split, mis-weighted rerank	pattern_query_parsing_split.md · rerankers.md
Citations missing or unstable	schema not enforced, formatter renamed fields	retrieval-traceability.md · data-contracts.md
Answers flip between runs	prompt header reordering or variance	context-drift.md · rerankers.md

Rebuild order when numbers stay bad

Follow the store-agnostic sequence and re-measure after each step. Open: Retrieval Playbook

Lock analyzer and distance metric
Re-chunk with anchor checklist
Re-embed with a single family and normalization
Add deterministic reranker and stabilize order
Tighten data contracts and traceability
Evaluate with the gold set and ΔS probes Open: retrieval_eval_recipes.md

Ops monitors to keep on

Index readiness fence and version hash Open: bootstrap-ordering.md
Live ΔS and λ alerts on long windows Open: ops live monitoring
Regression gate for coverage and ΔS Open: eval precision and recall

🔗 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.

12 KiB Raw Blame History Unescape Escape