🧭 Quick Return to Map
You are in a sub-page of Language.
To reorient, go back here:
- Language — multilingual processing and semantic alignment
- 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.
Stabilize hybrid retrieval across languages and scripts. Lock bm25 to dense weights, keep analyzers and tokenizers aligned, and verify with ΔS, λ, and coverage targets.
- Visual map and recovery → rag-architecture-and-recovery.md
- End to end retrieval knobs → retrieval-playbook.md
- Traceability schema → retrieval-traceability.md
- Contract the payload → data-contracts.md
- Tokenizer variance → tokenizer_mismatch.md
- Mixed scripts → script_mixing.md
- Locale normalization → locale_drift.md
- Romanization rules → romanization_transliteration.md
- Proper nouns and aliases → proper_noun_aliases.md
- Language detection → query_language_detection.md
- Analyzer routing → query_routing_and_analyzers.md
- ΔS(question, retrieved) ≤ 0.45 on three paraphrases and two seeds
- Coverage of target section ≥ 0.70
- λ convergent when switching between native script and romanized forms
- Rank@k is stable across analyzer choices for the same corpus and query set
- Reranker never flips a correct snippet out of top-k after hybrid fusion
| Symptom | Likely cause | Open this |
|---|---|---|
| High recall but unstable rank across runs | hybrid weight drift or nondeterministic tie handling | retrieval-playbook.md |
| Good dense hits, poor lexical for CJK | analyzer mismatch or missing bigram field | script_mixing.md |
| Romanized queries fail while native works | alias view absent in fusion | romanization_transliteration.md |
| JSON mode breaks after reranking | tokenizer profile differs between stages | tokenizer_mismatch.md |
| Brand equals common word in second language | alias collision and unscoped synonyms | proper_noun_aliases.md |
Use these as starting points then tune by ΔS and coverage. Always log weights and determinize tie rules.
- Analyzer: locale aware Latin
- Start weights: bm25 0.4, dense 0.6
- Reranker: on for top 100 → 20
- Tie rule: lexical score as primary tie breaker
- Analyzer: CJK bigram or language specific
- Start weights: bm25 0.55, dense 0.45
- Reranker: on, but keep citations from native script fields
- Add alias views: romaji or pinyin for entities only
- Analyzer: RTL with width and digit normalization
- Start weights: bm25 0.5, dense 0.5
- Reranker: on, strict JSON schema in later steps
- Analyzer: script aware with accent controls
- Start weights: bm25 0.45, dense 0.55
- Reranker: on, cross check aliases if brands exist
- Dual route: native analyzer plus romanized alias probe
- Start weights: bm25 0.5, dense 0.5
- Determinize the fusion and log both analyzer paths
- Fix seeds for dense stage and for reranker.
- Bucketize scores before fusion to avoid floating noise.
- Set a stable tie breaker: lexical score, then doc id.
- Cap max per-section to avoid one section flooding top-k.
- Fuse fields with the same analyzer class only.
- Keep the fusion function constant across languages inside the same pipeline.
A. Fusion function sketch
def fuse(bm25, dense, w_lex=0.5, w_vec=0.5):
# bm25 and dense are lists of (doc_id, score)
import math
L = {d: s for d, s in bm25}
V = {d: s for d, s in dense}
docs = set(L) | set(V)
out = []
for d in docs:
l = L.get(d, 0.0)
v = V.get(d, 0.0)
# bucketize to stabilize small drifts
lb = round(l, 3)
vb = round(v, 3)
score = w_lex * lb + w_vec * vb
out.append((d, score, lb, vb))
# stable sort: score desc, then lexical bucket, then doc id
out.sort(key=lambda x: (-x[1], -x[2], x[0]))
return outB. Rerank gate
def rerank_gate(items, target_k=20):
# never drop below k if citations exist and ΔS is already ≤ 0.45
keep = []
for d, score, lb, vb, meta in items:
if meta.get("has_citation") and meta.get("delta_s", 1.0) <= 0.45:
meta["protected"]=True
keep.append((d, score, meta))
return keep[:target_k]C. Trace fields to log
{
"fusion": {"w_bm25": 0.55, "w_dense": 0.45, "tie": "lex,doc_id"},
"analyzer_id": "store:cjk",
"tokenizer_profile": "retriever:cjk",
"alias_views": ["name_romaji","name_pinyin"],
"ΔS": 0.41,
"λ_state": "<>",
"coverage": 0.73
}
- Use bilingual sets from code_switching_eval.md.
- For each query, add a romanized twin and at least one brand alias variant.
- Report ΔS, coverage, λ on three paraphrases and two seeds.
- Accept only if: ΔS ≤ 0.45, coverage ≥ 0.70, λ convergent, no harmful reranker flips.
- Keep per language family the same reranker window size.
- ΔS stays ≥ 0.60 even after fusion tuning → revisit analyzer routing and re-chunk, open query_routing_and_analyzers.md and retrieval-playbook.md.
- Instability appears only in long chains → treat as reasoning drift, open logic-collapse.md and add a BBCR bridge.
You have TXTOS and WFGY Problem Map loaded.
Task:
1) For {query, lang, script}, choose hybrid weights deterministically.
2) Run cite-then-explain. Protect snippets that already meet ΔS ≤ 0.45.
3) Return a JSON trace:
{ "w_bm25": 0.xx, "w_dense": 0.xx, "ΔS": 0.xx, "coverage": 0.xx, "λ_state": "→|←|<>|×", "protected_ids": [...] }
Keep it auditable and short.
| 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 |
| Layer | Page | What it’s for |
|---|---|---|
| ⭐ Proof | WFGY Recognition Map | External citations, integrations, and ecosystem proof |
| ⚙️ Engine | WFGY 1.0 | Original PDF tension engine and early logic sketch (legacy reference) |
| ⚙️ Engine | WFGY 2.0 | Production tension kernel for RAG and agent systems |
| ⚙️ Engine | WFGY 3.0 | TXT based Singularity tension engine (131 S class set) |
| 🗺️ Map | Problem Map 1.0 | Flagship 16 problem RAG failure taxonomy and fix map |
| 🗺️ Map | Problem Map 2.0 | Global Debug Card for RAG and agent pipeline diagnosis |
| 🗺️ Map | Problem Map 3.0 | Global AI troubleshooting atlas and failure pattern map |
| 🧰 App | TXT OS | .txt semantic OS with fast bootstrap |
| 🧰 App | Blah Blah Blah | Abstract and paradox Q&A built on TXT OS |
| 🧰 App | Blur Blur Blur | Text to image generation with semantic control |
| 🏡 Onboarding | Starter Village | Guided entry point for new users |
If this repository helped, starring it improves discovery so more builders can find the docs and tools.
