🏥 Quick Return to Emergency Room
You are in a specialist desk.
For full triage and doctors on duty, return here:
- 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 sub-room.
If you want full consultation and prescriptions, go back to the Emergency Room lobby.
A compact hub to stabilize cross-lingual retrieval and reasoning.
Use this folder when your corpus or queries include CJK, RTL, Indic, Cyrillic, accented Latin, or frequent code-switching. No infra change required.
| Page | What it solves | Typical symptom |
|---|---|---|
| tokenizer_mismatch.md | Locks tokenization and segmentation for CJK/Thai/Indic | High sim but low recall on CJK/Thai; broken tokens |
| script_mixing.md | One query carries mixed scripts and analyzers split | Mixed Latin+CJK queries under-recall or flip |
| locale_drift.md | Normalization for width/accents/variants (Hans↔Hant) | zh-Hans/zh-Hant never co-retrieve; accent variants miss |
| multilingual_guide.md | End-to-end recipes and acceptance targets | Unsure where drift comes from across languages |
| proper_noun_aliases.md | Alias shield for names, brands, products | Proper nouns oscillate across spellings |
| romanization_transliteration.md | Romanization pairs and transliteration consistency | Inconsistent transliteration causes misses |
| query_language_detection.md | Stable language detection contract | Detection flips per run; routing becomes random |
| query_routing_and_analyzers.md | Route analyzers per language + parity w/ index | Search vs index behave differently |
| hybrid_ranking_multilingual.md | Deterministic hybrid rerank across languages | Multilingual ranking unstable, hybrid < single |
| stopword_and_morphology_controls.md | Clamp stopwords/lemmatizers to protect meaning | Negations/particles vanish; unit words lost |
| fallback_translation_and_glossary_bridge.md | Controlled translation bridge with glossary | Local path ΔS stays high; glossary needed |
| code_switching_eval.md | Bilingual & code-switch eval sets + checks | Cannot prove multilingual stability before ship |
- High similarity yet wrong meaning on bilingual or mixed-script corpora
- Citations point to the wrong section after translating the question
- Hybrid retrievers underperform a single retriever across languages
- Index looks healthy while coverage stays low for non-Latin scripts
- Names flip between native, transliteration, and English aliases
- zh-Hans and zh-Hant never co-retrieve; Thai recall drops for no reason
- ΔS(question, retrieved) ≤ 0.45 across language variants
- Coverage ≥ 0.70 to the intended section after repair
- λ_observe convergent across 3 paraphrases and 2 seeds
- E_resonance flat on long windows that mix scripts
- Citation fields complete; alias noise does not leak into evidence
| Symptom | Likely cause | Open this |
|---|---|---|
| High similarity yet wrong meaning | Embedding not multilingual or pre-normalization mismatch | embedding-vs-semantic.md |
| Citations jump sections after translation | Snippet schema too loose | data-contracts.md · retrieval-traceability.md |
| zh-Hans and zh-Hant never co-retrieve | Variant mapping and width rules missing | locale_drift.md |
| Thai or CJK recall collapses | Tokenizer mismatch or missing segmenter | tokenizer_mismatch.md |
| Mixed Latin + CJK query under-recalls | Analyzer split across scripts | script_mixing.md |
| Hybrid worse than single | Query parsing split or mis-weighted rerank | patterns/pattern_query_parsing_split.md · rerankers.md |
| Proper nouns oscillate | Missing alias fields and entity shield | proper_noun_aliases.md |
| Transliteration inconsistency | Romanization rules not aligned | romanization_transliteration.md |
| Language detection drifts | Detection contract weak or unlocked | query_language_detection.md |
| Search vs index disagree | Analyzer routing error | query_routing_and_analyzers.md |
| Ranking unstable across languages | Mono-lingual reranker or unaligned features | hybrid_ranking_multilingual.md |
| Negations/particles vanish | Stopword or morphology too aggressive | stopword_and_morphology_controls.md |
| Persistent high ΔS on local path | Need glossary-backed translation bridge | fallback_translation_and_glossary_bridge.md |
-
Detect language
Emit stable language + confidence. If unstable, fix detection first.
→ query_language_detection.md -
Lock normalization and analyzers
Keep locale, width, accents, and segmentation identical on write/read.
→ locale_drift.md · query_routing_and_analyzers.md -
Protect entities and syntax
Alias fields and romanization pairs; clamp stopwords/morphology for negations and units.
→ proper_noun_aliases.md · romanization_transliteration.md · stopword_and_morphology_controls.md -
Stabilize ranking
Use multilingual or dual-track rerank with deterministic ordering.
→ hybrid_ranking_multilingual.md -
Translation bridge only if needed
Pair with a glossary and keep native path as default.
→ fallback_translation_and_glossary_bridge.md -
Verify
With bilingual & code-switch sets confirm ΔS ≤ 0.45, Coverage ≥ 0.70, λ convergent.
→ code_switching_eval.md
- Normalize the same way for corpus and queries before storing vectors
- CJK/Thai require segmentation or bigrams; keep entity fields as keyword
- If no multilingual embeddings, add a lexical sidecar and align features with a deterministic rerank
Got it — here’s the English FAQ version for the Language & Multilingual · Global Fix Map README. It follows the same style and clarity as the Chinese one, but rewritten in English for new users.
Q1. Why does a bilingual or mixed query look similar but hit the wrong section?
A1. Most often the index and query use different analyzers or normalization steps, or CJK/Thai segmentation was never applied. Always lock the same normalization (width, accents, casing, segmentation) for both sides, then rebuild the index.
Open: tokenizer_mismatch.md · query_routing_and_analyzers.md
Q2. Why do zh-Hans and zh-Hant never co-retrieve?
A2. Variant and width rules are missing. Apply Unicode normalization, full/half-width mapping, and variant mapping before indexing.
Open: locale_drift.md
Q3. After translating the question into English, citations jump to the wrong section.
A3. The citation schema is too loose, missing fields like section_id and offsets. Enforce snippet contracts and cite-then-explain.
Open: data-contracts.md · retrieval-traceability.md
Q4. Why does Thai or Japanese recall fluctuate a lot?
A4. Classic tokenizer mismatch. Ensure index and query share the same segmenter; if not, use bigram or hybrid segmentation.
Open: tokenizer_mismatch.md
Q5. Why do mixed Latin + CJK queries under-recall?
A5. The analyzer splits into two routes and weights unevenly. Script-aware splitting or fixed routing is needed.
Open: script_mixing.md · query_routing_and_analyzers.md
Q6. Why do proper nouns oscillate between native, romanized, and English aliases?
A6. Alias fields and romanization tables are missing. Add aliases and protect them with keyword fields.
Open: proper_noun_aliases.md · romanization_transliteration.md
Q7. Why does multilingual reranking give different orderings each run?
A7. You are using a monolingual reranker or unaligned features. Switch to a multilingual reranker or dual-track (lexical+vector) with deterministic tie-breaks.
Open: hybrid_ranking_multilingual.md
Q8. Should I enable translation bridging from the start?
A8. No. Always try the native language path first. Only enable when ΔS stays above 0.45 over time, and always with glossaries.
Open: fallback_translation_and_glossary_bridge.md
Q9. Why do negations or particles disappear, flipping the meaning?
A9. Stopword or morphology rules are too aggressive. Protect negations, units, and structural particles.
Open: stopword_and_morphology_controls.md
Q10. Why does language detection keep flipping and causing misrouting?
A10. The detection contract isn’t locked, or samples are too short. Set stable model, sample length, confidence threshold, and fallback paths.
Open: query_language_detection.md
Q11. Metrics look fine but recall for non-Latin languages stays low.
A11. First check normalization and segmentation, then verify aliases/romanization and multilingual rerank alignment. Add code-switch eval sets for validation.
Open: multilingual_guide.md · code_switching_eval.md
Q12. What is the minimum acceptance test?
A12. Run bilingual and code-switch eval sets. Confirm all:
- ΔS(question, retrieved) ≤ 0.45
- Coverage ≥ 0.70
- λ convergent.
If not, debug in order: detection → normalization → entity protection → rerank → translation bridge.
Q13. Is there a ready-to-paste diagnostic prompt?
A13. Yes. Use the following inside your LLM:
You have TXTOS and the WFGY Problem Map loaded.
Task:
- Given a bilingual question Q, measure ΔS(Q, retrieved) and λ across 3 paraphrases.
- Verify index/query normalization (width, accents, casing, segmentation).
- Enforce cite-then-explain. Protect entities with alias/romanization.
- If ΔS ≥ 0.60 or λ flips, output minimal structural fix until ΔS ≤ 0.45, Coverage ≥ 0.70.
Return JSON:
{ "citations":[...], "ΔS":0.xx, "λ_state":"<>|→|←|×", "coverage":0.xx, "next_fix":"..." }Q14. If I want to change the least, what’s the fix priority? A14. 1) Lock language detection contract 2) Lock normalization and analyzers 3) Add aliases/romanization 4) Multilingual rerank 5) Only then enable translation bridge.
Q15. Accuracy improved, but rankings across languages still flip occasionally. A15. Add stable sort keys and fixed weight tables. Inject language features into rerankers and set deterministic tie-break rules. Open: hybrid_ranking_multilingual.md
| 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.
