6.8 KiB
Multilingual Guide — Guardrails and Fix Patterns
🧭 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.
A compact field guide to stabilize multilingual RAG across CJK, RTL, mixed scripts, and locale drift. Use this page to check symptoms, apply structural fixes, and verify with measurable targets.
Open these first
- Visual recovery map: RAG Architecture & Recovery
- Tokenization vs semantics: Embedding ≠ Semantic
- OCR parsing in multilingual docs: OCR Parsing Checklist
- Chunk joins across scripts: Semantic Chunking Checklist
- Why this snippet: Retrieval Traceability
- Snippet schema enforcement: Data Contracts
Core acceptance
- ΔS(question, retrieved) ≤ 0.45 across at least 2 languages
- Coverage ≥ 0.70 for the target section in each language
- λ remains convergent across three paraphrases in mixed scripts
- E_resonance stays flat for long bilingual/RTL runs
Common multilingual failure modes
| Symptom | Likely cause | Open this |
|---|---|---|
| Retrieval drops snippets when query is in Chinese or Japanese | Tokenizer mismatch (no whitespace segmentation) | tokenizer_mismatch.md |
| Citations collapse when Arabic/Hebrew text mixes with English | Script directionality conflict | script_mixing.md |
| High similarity but meaning flips across locale | Locale analyzer mismatch (stemming / stopwords) | locale_drift.md |
| HyDE/BM25 retrieval different per language | Query expansion language bias | hyde_multilingual.md |
Fix in 60 seconds
-
Probe with ΔS Run the same question in English and one target language. If ΔS differs by >0.15, suspect tokenization or analyzer mismatch.
-
Apply λ_observe Paraphrase the query three ways in the non-English language. If λ diverges, enforce schema lock and re-index with language-specific analyzers.
-
Structural repair
- Tokenizer mismatch → tokenizer_mismatch.md
- Script mixing issues → script_mixing.md
- Locale drift → locale_drift.md
- HyDE instability → hyde_multilingual.md
Diagnostic checklist
- Tokenizer: verify segmentation strategy (whitespace vs character-level)
- Analyzer: confirm stemming and stopword lists match query language
- Scripts: normalize Unicode, check RTL/LTR flags
- Locale drift: run same snippet under two locales, compare ΔS
- Hybrid retriever: ensure rerankers operate on normalized embeddings
🔗 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 | Canonical framework entry point | View |
| Problem Map | Diagnostic map and navigation hub | View |
| Tension Universe Experiments | MVP experiment field | View |
| Recognition | Where WFGY is referenced or adopted | View |
| AI Guide | Anti-hallucination reading protocol for tools | View |
If this repository helps, starring it improves discovery for other builders.
