feat: engraph v2.0 — hybrid search, smart chunking, vault profiles

.gitignore

@@ -2,3 +2,4 @@
 *.swp
 *.swo
 .DS_Store
+.worktrees/

CLAUDE.md

@@ -4,28 +4,35 @@ Local semantic search CLI for Obsidian vaults. Rust, MIT licensed.
 
 ## Architecture
 
-Single binary with 7 modules behind a lib crate:
-
-- `config.rs` — loads `~/.engraph/config.toml`, merges CLI args, provides `data_dir()`
-- `chunker.rs` — splits markdown by `##` headings, strips YAML frontmatter, extracts tags. `split_oversized_chunks()` handles token-aware sub-splitting with overlap
-- `embedder.rs` — downloads and runs `all-MiniLM-L6-v2` ONNX model (384-dim). SHA256-verified on download. Uses `ort` for inference, `tokenizers` for tokenization
-- `store.rs` — SQLite persistence. Tables: `meta`, `files`, `chunks` (with vector BLOBs), `tombstones`. Handles incremental diffing via content hashes
+Single binary with 11 modules behind a lib crate:
+
+- `config.rs` — loads `~/.engraph/config.toml` and `vault.toml`, merges CLI args, provides `data_dir()`
+- `chunker.rs` — smart chunking with break-point scoring algorithm. Finds optimal split points considering headings, code fences, blank lines, and thematic breaks. `split_oversized_chunks()` handles token-aware secondary splitting with overlap
+- `docid.rs` — deterministic 6-char hex IDs for files (SHA-256 of path, truncated). Shown in search results for quick reference
+- `embedder.rs` — downloads and runs `all-MiniLM-L6-v2` ONNX model (384-dim). SHA256-verified on download. Uses `ort` for inference, `tokenizers` for tokenization. Implements `ModelBackend` trait
+- `model.rs` — pluggable `ModelBackend` trait, model registry, and `parse_model_spec()`. Enables future model swapping without changing consumer code
+- `fts.rs` — FTS5 full-text search support. Re-exports `FtsResult` from store. BM25-ranked keyword search
+- `fusion.rs` — Reciprocal Rank Fusion (RRF) engine. Merges semantic + FTS5 results. Supports lane weighting and `--explain` output
+- `profile.rs` — vault profile detection. Auto-detects PARA/Folders/Flat structure, vault type (Obsidian/Logseq/Plain), wikilinks, frontmatter, tags. Writes/loads `vault.toml`
+- `store.rs` — SQLite persistence. Tables: `meta`, `files` (with docid), `chunks` (with vector BLOBs), `chunks_fts` (FTS5 virtual table), `tombstones`. Handles incremental diffing via content hashes
 - `hnsw.rs` — thin wrapper around `hnsw_rs`. **Important:** `hnsw_rs` does not support inserting after `load_hnsw()`. The index is rebuilt from vectors stored in SQLite on every index run
-- `indexer.rs` — orchestrates vault walking (via `ignore` crate for `.gitignore` support), diffing, chunking, embedding (Rayon for parallel chunking, serial embedding since `Embedder` is not `Send`), and serial writes to store + HNSW
-- `search.rs` — embeds query, searches HNSW with tombstone filtering, formats results (human + JSON). Also handles `status` formatting
+- `indexer.rs` — orchestrates vault walking (via `ignore` crate for `.gitignore` support), diffing, chunking, embedding (Rayon for parallel chunking, serial embedding since `Embedder` is not `Send`), and serial writes to store + HNSW + FTS5
 
-`main.rs` is a thin clap CLI that wires the modules together.
+`main.rs` is a thin clap CLI that wires the modules together. Subcommands: `index`, `search` (with `--explain`), `status`, `clear`, `init`, `configure`, `models`.
 
 ## Key patterns
 
-- **Incremental indexing:** `diff_vault()` compares file content hashes in SQLite against disk. Changed files have their old chunks deleted (cascade), then are re-embedded as new
+- **Hybrid search:** Queries run through two lanes — semantic (HNSW embeddings) and keyword (FTS5 BM25). Results are fused via Reciprocal Rank Fusion (RRF) with configurable lane weights
+- **Smart chunking:** Break-point scoring algorithm assigns scores to potential split points (headings 50-100, code fences 80, thematic breaks 60, blank lines 20). Chunks split at the highest-scored break point near the token target. Code fence protection prevents splitting inside code blocks
+- **Incremental indexing:** `diff_vault()` compares file content hashes in SQLite against disk. Changed files have their old chunks deleted (cascade), then are re-embedded as new. FTS5 entries are cleaned up alongside vector entries
 - **HNSW rebuild on every run:** Vectors are stored as BLOBs in the `chunks` table. After SQLite is updated, the full HNSW index is rebuilt from `store.get_all_vectors()`. This is necessary because `hnsw_rs` doesn't support append-after-load
-- **Vector IDs:** Assigned sequentially, stored in both SQLite and HNSW. `next_vector_id` is derived from `MAX(vector_id)` in SQLite
-- **Tombstones:** Exist in the schema but are largely unused now that we rebuild HNSW each run. Kept for future use if switching to a vector store that supports deletion
+- **Docids:** Each file gets a deterministic 6-char hex ID (SHA-256 of relative path). Displayed in search results for quick reference
+- **Vault profiles:** `engraph init` auto-detects vault structure and writes `vault.toml`
+- **Pluggable models:** `ModelBackend` trait enables future model swapping. Current implementation uses ONNX all-MiniLM-L6-v2
 
 ## Data directory
 
-`~/.engraph/` — hardcoded via `Config::data_dir()` (uses `dirs::home_dir()`). Contains `engraph.db` (SQLite), `hnsw/` (index files), `models/` (ONNX model + tokenizer).
+`~/.engraph/` — hardcoded via `Config::data_dir()` (uses `dirs::home_dir()`). Contains `engraph.db` (SQLite with FTS5), `hnsw/` (index files), `models/` (ONNX model + tokenizer), `vault.toml` (vault profile), `config.toml` (user config).
 
 Single vault only. Re-indexing a different vault path triggers a confirmation prompt.
 
@@ -35,10 +42,11 @@ Single vault only. Re-indexing a different vault path triggers a confirmation pr
 - `hnsw_rs` (0.3) — pure Rust HNSW. `Box::leak` used in `load()` to satisfy `'static` lifetime on the loaded index. Read-only after load
 - `tokenizers` (0.22) — HuggingFace tokenizer. Needs `fancy-regex` feature
 - `ignore` (0.4) — vault walking with automatic `.gitignore` support
+- `rusqlite` (0.32) — bundled SQLite with FTS5 support
 
 ## Testing
 
-- Unit tests in each module (`cargo test --lib`) — 44 tests, no network required
+- Unit tests in each module (`cargo test --lib`) — 91 tests, no network required
 - 1 ignored smoke test (`test_embed_smoke`) — downloads ONNX model, verifies embedding
 - Integration tests (`cargo test --test integration -- --ignored`) — 8 tests, require model download. Use `tempfile` for isolated data dirs
 

Cargo.toml

@@ -1,6 +1,6 @@
 [package]
 name = "engraph"
-version = "0.1.0"
+version = "0.2.0"
 edition = "2024"
 description = "Local semantic search for Obsidian vaults"
 license = "MIT"