Korean-Specialized RAG
Korean RAG quality comes from document structure, morphological variation, organizational terminology, and citation verification—not vector search alone.
Design the path from Korean document ingestion through structure-preserving chunking, hybrid retrieval, reranking, evidence, and access control.
01
Problem definition
Spacing, endings, abbreviations, and Korean-English mixing destabilize recall and ranking in embedding-only retrieval.
A relevant sentence cannot support work if tables, footnotes, and revision state are lost.
- Korean regulations, manuals, or reports must ground answers
- Document authority and citations must be audited per answer
- Sources change frequently but revision state cannot be managed
- Structured queries can answer the data exactly without retrieval
Korean-Specialized RAG: system plate
- 01
Stage 1
Ingest HWP, HWPX, and PDF while preserving headings, tables, footnotes, and revision metadata.
- 02
Stage 2
Combine sparse and dense retrieval with morphology, initial consonants, and organizational synonyms.
- 03
Stage 3
Tune candidate generation and reranking by query type against held-out questions.
- 04
Stage 4
Validate sentence/table-cell citations, authority filters, and abstention as an answer contract.
- The workflow begins with Ingest HWP, HWPX, and PDF while preserving headings, tables, footnotes, and revision metadata..
- It reaches an acceptance decision through Evidence recall@k and ranking quality.
03
Design method
Korean RAG quality comes from document structure, morphological variation, organizational terminology, and citation verification—not vector search alone.
- 01
Stage 1
Ingest HWP, HWPX, and PDF while preserving headings, tables, footnotes, and revision metadata.
- 02
Stage 2
Combine sparse and dense retrieval with morphology, initial consonants, and organizational synonyms.
- 03
Stage 3
Tune candidate generation and reranking by query type against held-out questions.
- 04
Stage 4
Validate sentence/table-cell citations, authority filters, and abstention as an answer contract.
04
Application scenarios
Korean regulations, manuals, or reports must ground answers
Spacing, endings, abbreviations, and Korean-English mixing destabilize recall and ranking in embedding-only retrieval.
- APPROACH
- Ingest HWP, HWPX, and PDF while preserving headings, tables, footnotes, and revision metadata.
- BOUNDARY
- Retrieval hits are not equated with answer factuality.
Document authority and citations must be audited per answer
A relevant sentence cannot support work if tables, footnotes, and revision state are lost.
- APPROACH
- Combine sparse and dense retrieval with morphology, initial consonants, and organizational synonyms.
- BOUNDARY
- Model memory must not bypass document authority for unauthorized users.
05
Design choices
| Decision | Gain | Cost | Watch |
|---|---|---|---|
| Korean regulations, manuals, or reports must ground answers | Ingest HWP, HWPX, and PDF while preserving headings, tables, footnotes, and revision metadata. | Retrieval hits are not equated with answer factuality. | Evidence recall@k and ranking quality |
| Document authority and citations must be audited per answer | Combine sparse and dense retrieval with morphology, initial consonants, and organizational synonyms. | Model memory must not bypass document authority for unauthorized users. | Citation correctness and completeness |
Korean-Specialized RAG: system plate
| Item | Method | Evidence | Boundary |
|---|---|---|---|
| Layer 1 | Ingest HWP, HWPX, and PDF while preserving headings, tables, footnotes, and revision metadata. | Evidence recall@k and ranking quality | Retrieval hits are not equated with answer factuality. |
| Layer 2 | Combine sparse and dense retrieval with morphology, initial consonants, and organizational synonyms. | Citation correctness and completeness | Model memory must not bypass document authority for unauthorized users. |
| Layer 3 | Tune candidate generation and reranking by query type against held-out questions. | Authority leakage and abstention accuracy | Retrieval hits are not equated with answer factuality. |
- The workflow begins with Ingest HWP, HWPX, and PDF while preserving headings, tables, footnotes, and revision metadata..
- It reaches an acceptance decision through Evidence recall@k and ranking quality.
07
Validation plan
| Measure | Method | Pass condition | Caveat |
|---|---|---|---|
| Evidence recall@k and ranking quality | Ingest HWP, HWPX, and PDF while preserving headings, tables, footnotes, and revision metadata. | Repeated runs satisfy the acceptance threshold agreed during discovery | Retrieval hits are not equated with answer factuality. |
| Citation correctness and completeness | Combine sparse and dense retrieval with morphology, initial consonants, and organizational synonyms. | Repeated runs satisfy the acceptance threshold agreed during discovery | — |
| Authority leakage and abstention accuracy | Tune candidate generation and reranking by query type against held-out questions. | Repeated runs satisfy the acceptance threshold agreed during discovery | — |
08
Constraints and failure conditions
Sources change frequently but revision state cannot be managed
Retrieval hits are not equated with answer factuality.
Structured queries can answer the data exactly without retrieval
Model memory must not bypass document authority for unauthorized users.
09
Engagement model
- 01
Diagnosis
PattyAnalyze the current system and its failure signals.
ClientProvide representative work, data boundaries, and operating constraints.
Evidence recall@k and ranking quality - 02
Design
PattyCombine sparse and dense retrieval with morphology, initial consonants, and organizational synonyms.
ClientConfirm owners and acceptance criteria.
Citation correctness and completeness - 03
Validation
PattyTune candidate generation and reranking by query type against held-out questions.
ClientMake the production-transition or stop decision.
Authority leakage and abstention accuracy
10
Durable deliverables
- Korean-Specialized RAG decision record
- Korean RAG quality comes from document structure, morphological variation, organizational terminology, and citation verification—not vector search alone.Client-owned · Patty-reviewed
- Validation harness and acceptance criteria
- Evidence recall@k and ranking quality · Citation correctness and completeness · Authority leakage and abstention accuracyJointly maintained
- Operations and recovery runbook
- Retrieval hits are not equated with answer factuality. · Model memory must not bypass document authority for unauthorized users.Operating-team owned
11
Terminology
- Korean-Specialized RAG
- Design the path from Korean document ingestion through structure-preserving chunking, hybrid retrieval, reranking, evidence, and access control.
- Acceptance criterion
- Evidence recall@k and ranking quality
- Operating boundary
- Retrieval hits are not equated with answer factuality.
REFERENCES
References and primary material
- BEIR
Primary material for the method and terminology.
- NIST AI RMF: Map
Primary material for the method and terminology.