RAG & Knowledge Base

SalmAlm includes a zero-dependency RAG (Retrieval-Augmented Generation) system — no external vector DB, no embeddings API. Everything runs locally in SQLite.

How It Works

User Query → Tokenize → BM25 + TF-IDF Hybrid Search → Top-K Chunks → Inject into Prompt

1. Indexing: Files are chunked (5 lines, 2-line overlap) and tokenized with Korean jamo decomposition + English stemming
2. Search: Hybrid BM25 (text weight 0.3) + TF-IDF cosine similarity (vector weight 0.7)
3. Injection: Top results are prepended to the system prompt as context

What Gets Indexed

Source	Auto-indexed	Tool
Memory files (~/SalmAlm/memory/)	✅	memory_write
Notes	✅	note
Saved links	✅	save_link
Workspace files	On demand	file_index
Session transcripts	Optional	Config toggle

Configuration

Edit ~/SalmAlm/rag.json:

{
  "hybrid": {
    "enabled": true,
    "vectorWeight": 0.7,
    "textWeight": 0.3
  },
  "sessionIndexing": {
    "enabled": false,
    "retentionDays": 30
  },
  "chunkSize": 5,
  "chunkOverlap": 2,
  "reindexInterval": 120
}

Korean Language Support

SalmAlm's RAG has first-class Korean support:

• Jamo decomposition — decomposes Hangul into consonants/vowels for fuzzy matching
• Korean stemming — strips common suffixes (은/는/이/가/을/를/에서/으로 etc.)
• Synonym expansion — 검색→찾기/탐색, 오류→에러/버그 etc.
• Stop word filtering — removes Korean particles and common English stop words

Web UI

Access the RAG panel at Settings → RAG & Knowledge to:

• View index statistics (document count, chunk count, last indexed)
• Trigger manual reindex
• Configure hybrid search weights
• Enable/disable session indexing

Related Tools

• rag_search — Direct semantic search
• memory_write / memory_search — Long-term memory CRUD
• note — Quick notes (auto-indexed)
• save_link — Save URLs with auto-fetched content
• file_index — Index workspace files on demand