Architecture Overview

ContextWeaver is organized into four layers: interface, indexing, storage, and search.

Interface layer

File	Responsibility
`src/index.ts`	CLI entrypoint; registers `init/index/watch/search/mcp/migrate/stats` commands
`src/mcp/server.ts`	MCP server; registers tools and handles protocol calls
`src/cli/mirrorCommands.ts`	Mirrors selected MCP tools as CLI commands

The indexing layer reads source files, filters low-value files, computes hashes, chunks code, and writes indexes.

File	Responsibility
`src/scanner/crawler.ts`	Traverses the filesystem with `fdir`
`src/scanner/filter.ts`	Applies default ignore rules and `IGNORE_PATTERNS`
`src/scanner/processor.ts`	Reads files, detects encoding, and computes hashes
`src/chunking/SemanticSplitter.ts`	AST semantic chunking and fallback line chunking
`src/indexer/index.ts`	Batch embedding, LanceDB writes, FTS writes, and SQLite marks

ContextWeaver uses two storage systems:

SQLite: file metadata, full content, FTS5, statistics, migration state, and outbox
LanceDB: vectors and locating metadata

Since v1.4.0, source content only lives in SQLite files.content; LanceDB no longer stores display_code or vector_text.

The search layer builds final context packs from user queries.

File	Responsibility
`src/search/SearchService.ts`	Hybrid retrieval, RRF, rerank, Smart TopK, and stats recording
`src/search/QueryCache.ts`	Per-project LRU query cache
`src/search/GraphExpander.ts`	E1/E2/E3 context expansion
`src/search/ContextPacker.ts`	Segment merging and token-budget control
`src/search/ChunkContentLoader.ts`	Slices content from `files.content`
`src/search/fts.ts`	File-level and chunk-level FTS search