ripvec

Cacheless semantic code + document search that ties or beats transformer baselines. One binary, 19 grammars, three retrieval engines, zero setup.

ripvec finds code and documents by meaning, provides structural code intelligence across every language it knows, and ranks results by how important each file is in your project. The default engine runs CPU-only, holds no on-disk index, and matches or exceeds ModernBERT-class transformers on our benchmark matrix across code and prose. Transformer engines remain available, opt-in, for users who want a persistent index and the best coherent top-K on long-form narrative.

$ ripvec "retry logic with exponential backoff" ~/src/my-project

 1. retry_handler.rs:42-78                                        [0.91]
    pub async fn with_retry<F, T>(f: F, max_attempts: u32) -> Result<T>
    where F: Fn() -> Future<Output = Result<T>> {
        let mut delay = Duration::from_millis(100);
        for attempt in 0..max_attempts {
            match f().await {
                Ok(v) => return Ok(v),
                Err(e) if attempt < max_attempts - 1 => {
                    sleep(delay).await;
                    delay *= 2;  // exponential backoff
    ...

 2. http_client.rs:156-189                                        [0.84]
    impl HttpClient {
        async fn request_with_backoff(&self, req: Request) -> Response {
    ...

The function is called with_retry, the variable is delay. "exponential backoff" appears nowhere in the source. grep can't find this. ripvec can, because it embeds both your query and the code into the same vector space, fuses semantic scores with path-enriched BM25, layers a structural-importance signal from a PageRank percentile boost, and reranks the top candidates through a cross-encoder.

When to use what

ripvec has three interfaces. Here's when each one matters:

The MCP server gives AI agents 8 semantic + structural tools plus 9 LSP tools. The LSP server gives editors structural intelligence (outlines, go-to-definition, syntax diagnostics) for all 19 languages from one binary. The CLI is for humans. Same binary for all three.

If you're using Claude Code, install the plugin. It sets up both MCP and LSP automatically; Claude will use search_code when you ask conceptual questions and the LSP for symbol navigation.

Engines

Three retrieval engines share the same CLI/MCP/LSP surface. Pick at runtime with --model:

graph TB
    Q["Query"] --> S["Shared surface<br/>CLI / MCP / LSP"]
    S --> R["--model ripvec<br/>(default)"]
    S --> M["--model modernbert"]
    S --> B["--model bert"]
    R --> RP["Model2Vec 32M bi-encoder<br/>+ path-enriched BM25<br/>+ PageRank percentile boost<br/>+ MiniLM-L-12 cross-encoder rerank<br/>= in-memory only"]
    M --> MP["ModernBERT 768-dim transformer<br/>+ BM25 + PageRank<br/>= persistent on-disk index"]
    B --> BP["BGE-small 384-dim transformer<br/>+ BM25 + PageRank<br/>= persistent on-disk index"]

The MCP daemon picks the engine at startup via RIPVEC_MCP_ENGINE (defaults to ripvec); the transformer engines require building ripvec-mcp with --features legacy-transformer-mcp. The CLI accepts --model per-invocation with no build-time gating.

Quality

Measured target-hit count in top-3 results across four corpora, 10 queries each. A "target hit" is a result from the file or book a human evaluator considered correct for the query. Per corpus the maximum is 30 (10 queries × top-3 each); the TOTAL column has a maximum of 120.

The ripvec engine ties or beats ModernBERT on top-3 across this matrix. ModernBERT retains an edge on top-10 coherence for long-form narrative prose: on 3 of 10 Gutenberg queries, every BERT result is from the right book where ripvec spreads positions 2-10 across noise. That gap is real and is the structural cost of static-embedding retrieval. Most code and document-search workloads care about top-3 far more than top-10 coherence.

The Tokio result is worth flagging. Under ModernBERT, 7/10 dev-flavored queries put test files in top-5 ahead of canonical implementations. The ripvec engine's PageRank percentile boost flips that: implementations rank above the tests that reference them, because tests have near-zero structural PageRank and implementations don't.

Workflow: orient, search, navigate

graph LR
    A["🗺️ Orient<br/>get_repo_map"] --> B["🔍 Search<br/>search_code"]
    B --> C["🧭 Navigate<br/>LSP operations"]
    C -->|"need more context"| B
    C -->|"found it"| D["✏️ Edit"]

Orient. get_repo_map returns a structural overview ranked by function-level importance. One tool call replaces 10+ sequential file reads. Start here when working on unfamiliar code.

Search. search_code "authentication middleware" finds implementations by meaning across all 19 languages simultaneously. Results are ranked by relevance and structural importance.

Navigate. LSP documentSymbol shows the file outline. goToDefinition jumps to the likely definition. findReferences shows usage sites. incomingCalls/outgoingCalls traces the call graph.

Semantic search

You describe behavior, ripvec finds the implementation:

Search modes: --mode hybrid (default, semantic + BM25 fusion), --mode semantic (pure vector similarity), --mode keyword (pure BM25). Hybrid is usually best.

Multi-language LSP

ripvec serves LSP from a single binary for all 19 grammars. No per-language server installs. It provides:

• documentSymbol: file outline (functions, fields, enum variants, constants, types, headings)
• workspaceSymbol: cross-language symbol search with PageRank boost
• goToDefinition: name-based resolution ranked by structural importance
• findReferences: usage sites via hybrid search + content filtering
• hover: scope chain, signature, enriched context
• publishDiagnostics: tree-sitter syntax error detection after every edit
• incomingCalls / outgoingCalls: function-level call graph

For languages with dedicated LSPs (Rust, Python, Go, TypeScript), ripvec runs alongside them. The dedicated server handles types, ripvec handles semantic search and cross-language features. For languages without dedicated LSPs (bash, HCL, Ruby, Kotlin, Swift, Scala), ripvec is the primary code intelligence.

JSON, YAML, TOML, and Markdown get structural outlines (keys, mappings, headings) and syntax diagnostics. Useful for navigating large config files, not comparable to language-aware intelligence.

Architecture: the ripvec engine

The default engine is a four-stage composite pipeline. Each stage uses a fast cheap-to-rebuild signal; together they outperform a single transformer on retrieval quality.

graph TB
    Q["Query"] --> EMB["Bi-encoder embed<br/>(Model2Vec 32M)"]
    Q --> BM["BM25 score<br/>(path-enriched tokenization)"]
    EMB --> SEM["Cosine similarity<br/>top-N candidates"]
    BM --> LEX["Lexical ranking<br/>top-N candidates"]
    SEM --> RRF["Reciprocal Rank Fusion<br/>(k=60)"]
    LEX --> RRF
    RRF --> PR["× PageRank percentile boost<br/>(sigmoid curve, α=0.5)"]
    PR --> RR["Cross-encoder rerank<br/>(ms-marco-MiniLM-L-12-v2)<br/>top-100 candidates"]
    RR --> OUT["Top-k results"]

Static bi-encoder retrieval (Model2Vec). The bi-encoder is a lookup-and-mean-pool over a pretrained 256-dim embedding table (minishlab/potion-base-32M). No transformer forward pass; encoding cost is dominated by memory bandwidth, not FLOPs. About 5ms per query on a single CPU thread; ~250K chunks per second when indexing in parallel.

Path-enriched BM25. Lexical scoring with a code-aware tokenizer that splits parseJsonConfig into [parse, json, config] and my_func_name into [my, func, name]. Chunk text is enriched with the file stem (doubled) and the last three directory components before tokenization, so a query like "session encoding" hits both content and sessions.py paths.

Reciprocal Rank Fusion. Combines the semantic and lexical rankings via Cormack et al.'s rank-based fusion (k=60). Handles the scale mismatch between cosine similarity and BM25 without tuning.

PageRank percentile boost. A structural-importance signal on top of relevance. See the next section.

Cross-encoder rerank. The top 100 candidates are re-scored by a cross-encoder (ms-marco-MiniLM-L-12-v2) that runs full attention across the concatenated query+document pair. Per-pair cost ~10ms on CPU; total ~1s for 100 candidates. Final scores are blended (0.7 × cross-encoder + 0.3 × bi-encoder) so when the cross-encoder is uncertain (sigmoid outputs clustered near 0.5), the bi-encoder's confident ordering is preserved.

Function-level PageRank

graph LR
    subgraph "Call Graph"
        A["main()"] --> B["handle_request()"]
        A --> C["init_db()"]
        B --> D["authenticate()"]
        B --> E["dispatch()"]
        D --> F["verify_token()"]
        E --> D
    end
    subgraph "PageRank"
        D2["authenticate() ★★★"]
        B2["handle_request() ★★"]
        E2["dispatch() ★"]
    end

ripvec extracts call expressions from every function body using tree-sitter, resolves callee names to definitions, and computes PageRank on the resulting call graph. Functions called by many others rank higher. authenticate() in the example above is more structurally important than dispatch() because more code depends on it.

The bi-encoder is structurally weaker than a transformer. Model2Vec doesn't model cross-token interactions and can't reliably distinguish a 1500-char canonical implementation from a 3-line example stub by dense similarity alone. Without a corrective signal, the engine ranks tests/hello_world.py competitively with src/auth/handler.py on a query like "register a route." PageRank carries the missing signal: implementations are imported by tests and callers; stubs are imported by nothing.

ripvec applies the structural prior as a sigmoid-on-percentile boost: boost(p) = 1 + α × sigmoid((p − 0.5) / s) where p is the file's PR percentile within the corpus, α=0.5 is the ceiling lift, and s=0.15 controls steepness.

Two design constraints fall out of this curve:

1. At-or-above-median PR gets a meaningfully different boost from low-PR. A median-importance impl with cosine 0.84 ends at 0.84 × 1.25 = 1.05; a near-zero-PR test with cosine 0.85 ends at 0.85 × 1.02 = 0.867. The impl flips above the test by ~21%, enough to reorder reliably when the bi-encoder is uncertain.
2. The ceiling caps centers-of-universe. A graph-root file at p=1.0 gets at most 1.5×. It can't dominate when the query genuinely matches a less-central file.

The boost is applied via a composable RankingLayer chain shared across CLI, MCP, and LSP code paths. Adding a new ranking signal (recency, file-saturation diversification) is a single new impl RankingLayer.

Architecture: transformer engines

ripvec retains two transformer engines for users who want a persistent on-disk index and the absolute coherent top-K on long-form prose. Both share the cache architecture, the BM25/RRF/PageRank ranking layers, and the GPU backends; they differ only in the embedding model.

Cache layout

graph TD
    subgraph "~/.cache/ripvec/<project_hash>/v3-modernbert/"
        M["manifest.json<br/>file entries + Merkle hashes"]
        L["manifest.lock<br/>advisory fd-lock"]
        subgraph "objects/ (content-addressed)"
            O1["ab/cdef12...<br/>(zstd-compressed FileCache)"]
            O2["3f/a891bc...<br/>(zstd-compressed FileCache)"]
        end
    end

Each file's chunks and embeddings are serialized into a FileCache object, compressed with zstd (~8x), and stored by blake3 content hash in a git-style xx/hash sharded object store. The manifest tracks metadata: mtime, size, content hash, chunk count per file, plus Merkle directory hashes.

The ripvec engine never builds any of this. It holds the in-memory index across an MCP session lifetime, drops it on reindex or process exit, and rebuilds on next query.

Change detection: two-level diff

graph TD
    F["File on disk"] --> M{"mtime + size<br/>match manifest?"}
    M -->|"yes"| SKIP["Unchanged<br/>(fast path, no I/O)"]
    M -->|"no"| HASH{"blake3 content hash<br/>matches manifest?"}
    HASH -->|"yes"| TOUCH["Touched but identical<br/>(heal mtime in manifest)"]
    HASH -->|"no"| DIRTY["Dirty → re-embed"]

Level 1 (mtime+size) is a stat call (microseconds). Level 2 (blake3 hash) reads the file but avoids re-embedding if content hasn't changed. After git clone (where all mtimes are wrong), the first run hashes everything but re-embeds nothing, then heals the manifest mtimes for fast-path on subsequent runs.

Serialization: two formats

Auto-detected on read via magic bytes: 0x42 0x43 = bitcode, otherwise rkyv. Both are zstd-compressed. Repo-level indices use bitcode so they can be committed to git and shared between x86 CI and ARM developer machines.

Concurrency and locking

sequenceDiagram
    participant MCP as MCP Server
    participant Watcher as File Watcher
    participant Lock as manifest.lock
    participant Cache as Object Store

    Note over MCP: Query arrives
    MCP->>Lock: acquire read lock
    MCP->>Cache: load objects
    Lock-->>MCP: release

    Note over Watcher: File change detected (2s debounce)
    Watcher->>Lock: acquire write lock
    Watcher->>Cache: re-embed dirty files
    Watcher->>Cache: write new objects
    Watcher->>Cache: save manifest + GC
    Lock-->>Watcher: release

The file watcher debounces for 2 seconds of quiet before triggering re-indexing. Advisory fd-lock on manifest.lock prevents readers from seeing a half-written manifest. Multiple readers can proceed concurrently; writers block all readers.

Garbage collection runs after each incremental update; unreferenced objects (from deleted or re-embedded files) are removed from the store.

Repo-level indexing

ripvec --model modernbert --index --repo-level "query"
git add .ripvec/ && git commit -m "add search index"

Creates .ripvec/config.toml (pins model + version) and .ripvec/cache/ (manifest + objects). Teammates who clone get instant search. The config is validated on load. If the model doesn't match the runtime model, ripvec falls back to the user-level cache with a warning.

Repo config can also exclude files from the index using .gitignore syntax:

[ignore]
patterns = [
  "*.jsonl",
  "*.md",
  "docs/generated/**",
  "!docs/README.md",
]

These patterns apply to CLI indexing, incremental cache diffing, MCP reindexing, and repo-map file discovery. The command-line --exclude-extensions=jsonl,md flag is useful for one-off extension filters.

Cache resolution

graph TD
    A["--cache-dir override"] -->|"highest priority"| R["Resolved cache dir"]
    B[".ripvec/config.toml<br/>(repo-local)"] -->|"if model matches"| R
    C["RIPVEC_CACHE env var"] --> R
    D["~/.cache/ripvec/<br/>(XDG default)"] -->|"lowest priority"| R

Embedding pipeline

graph LR
    subgraph "Stage 1: Chunk (rayon)"
        F["Files"] --> TS["Tree-sitter<br/>parse"]
        TS --> C["Semantic<br/>chunks"]
    end
    subgraph "Stage 2: Tokenize"
        C --> T["Tokenizer<br/>(BPE / WordPiece)"]
        T --> B["Padded<br/>batches"]
    end
    subgraph "Stage 3: Embed (GPU)"
        B --> FW["Forward pass<br/>(22 layers ModernBERT,<br/>12 layers BGE-small)"]
        FW --> P["Mean / CLS pool<br/>+ L2 norm"]
        P --> V["Embedding<br/>vectors"]
    end
    C -.->|"bounded channel<br/>backpressure"| T
    T -.->|"bounded channel<br/>backpressure"| FW

For large corpora (1000+ files), stages run concurrently as a streaming pipeline with bounded channels for backpressure. The GPU starts embedding after the first batch (~50ms), not after all files are chunked.

Driver / Architecture split

The core design insight for the transformer engines: the forward pass is written ONCE as a generic ModernBertArch<D: Driver>, and each backend implements the Driver trait with platform-specific operations. Same model, same math, different hardware.

graph TB
    subgraph "Architecture (written once)"
        FP["ModernBertArch&lt;D: Driver&gt;<br/>forward()"]
        FP --> L1["Layer 1: Attention + FFN"]
        L1 --> L2["Layer 2: Attention + FFN"]
        L2 --> LN["...22 layers..."]
        LN --> Pool["Mean pool + L2 norm"]
    end
    subgraph "Driver trait implementations"
        FP -.->|"D = Metal"| M["MetalDriver<br/>MPS GEMMs + custom MSL kernels"]
        FP -.->|"D = CUDA"| CU["CudaDriver<br/>cuBLAS tensor cores + NVRTC kernels"]
        FP -.->|"D = CPU"| CP["CpuDriver<br/>ndarray + Accelerate/OpenBLAS"]
        FP -.->|"D = MLX"| ML["MlxDriver<br/>lazy eval → auto-fused Metal"]
    end

What each backend actually does per layer

Each of the 22 ModernBERT layers runs attention + FFN. Here's how the same operations map to different hardware:

graph LR
    subgraph "Attention"
        LN1["LayerNorm"] --> QKV["QKV projection<br/>(GEMM)"]
        QKV --> PAD["Pad + Split"]
        PAD --> ROPE["RoPE rotation"]
        ROPE --> ATTN["Q @ K^T<br/>(batched GEMM)"]
        ATTN --> SM["Scale + Mask<br/>+ Softmax"]
        SM --> AV["Scores @ V<br/>(batched GEMM)"]
        AV --> UNPAD["Reshape + Unpad"]
        UNPAD --> OPROJ["Output proj<br/>(GEMM)"]
        OPROJ --> RES1["Residual add"]
    end
    subgraph "FFN"
        RES1 --> LN2["LayerNorm"]
        LN2 --> WI["Wi projection<br/>(GEMM)"]
        WI --> GEGLU["Split + GeGLU"]
        GEGLU --> WO["Wo projection<br/>(GEMM)"]
        WO --> RES2["Residual add"]
    end

Metal and CUDA have hand-written fused kernels for softmax, GeGLU, and attention reshape. These eliminate intermediate buffers and reduce memory bandwidth. MLX gets fusion automatically via lazy evaluation (the entire forward pass typically compiles to 2-3 Metal kernel dispatches). CPU uses explicit scalar loops for everything except GEMM.

Dual search index (transformer engines)

graph LR
    subgraph "HybridIndex"
        subgraph "SearchIndex (dense vectors)"
            EMB["embeddings<br/>(TurboQuant 4-bit compressed)"]
            EMB --> CS["Cosine similarity scan"]
        end
        subgraph "Bm25Index (tantivy)"
            TAN["Inverted index<br/>(code-aware tokenizer)"]
            TAN --> BM["BM25 scoring<br/>(name 3× / path 1.5× / body 1×)"]
        end
        CS --> RRF["RRF fusion (k=60)"]
        BM --> RRF
    end

The transformer BM25 index uses a code-aware tokenizer that splits parseJsonConfig into [parse, json, config] and my_func_name into [my, func, name]. Keyword search finds json config parser even if the function is named in camelCase. Function names are boosted 3x over body text.

TurboQuant compresses 768-dim vectors from 3KB to ~380 bytes (4-bit) with a rotation matrix for better quantization. This enables ~5x faster scanning for large indices while maintaining ranking quality through exact re-ranking of the top candidates.

Performance

Cacheless (ripvec engine, the default). Wall time for a single query, end-to-end including model load on cold start:

The MCP daemon holds the in-memory index for the session lifetime, so warm latency dominates after the first query. For sub-MCPs and agent fan-out where each spawn starts fresh, the cold-path numbers are what to budget against. Roughly 100× faster cold-path than ModernBERT cacheless (33s/79s on Gutenberg/Tokio respectively).

Indexed (transformer engines). Time to build the persistent index on first run; subsequent queries against the cached index are milliseconds.

Metal and CPU show similar throughput on M2 Max because macOS Accelerate routes BLAS operations through the AMX coprocessor regardless of backend. The Metal backend has headroom on larger batches and non-BLAS operations.

Memory. Ripvec engine: ~200 MB for a typical project (embedding table + chunks + BM25). Transformer engines: ~500 MB during embedding (model weights + batch buffers), ~100 MB for query-time.

Where CPU goes on the ripvec engine (linux/92K corpus, sampled).

The 38% sync floor is structural: rayon's par_iter join semantics require parking workers between stages. We've shipped what's worth shipping past that floor (mimalloc, hand-vectorized pool_ids, bounded-queue streaming pipeline, lasso term interning). Further compression would require restructuring around an async stage scheduler.

How it compares

ripvec is self-contained (no Ollama, no cloud, no per-language setup), runs locally, and combines search + LSP + structural ranking in one binary. The cacheless default fits sub-MCP / fan-out / fresh-worktree workflows where a persistent index isn't viable.

Install

Pre-built binaries (fastest)

cargo binstall ripvec ripvec-mcp

Requires cargo-binstall. Downloads a pre-built binary for your platform; no compilation.

From source

cargo install ripvec ripvec-mcp

For CUDA (Linux with NVIDIA GPU, transformer engines only):

cargo install ripvec ripvec-mcp --features cuda

To enable transformer engines on the MCP daemon:

cargo install ripvec-mcp --features legacy-transformer-mcp

(The default ripvec-mcp build ships only the ripvec engine. The CLI binary ripvec accepts all engines without feature gating.)

Claude Code plugin

claude plugin install ripvec@fnordpig-my-claude-plugins

The plugin auto-downloads the binary for your platform on first use and configures both MCP and LSP servers. It includes 3 skills (codebase orientation, semantic discovery, change impact analysis), 3 commands (/map, /find, /repo-index), and a code exploration agent. CUDA is auto-detected via nvidia-smi.

Platforms

Model weights download automatically on first run: ~33MB (potion-base-32M, default ripvec engine) or ~100MB (ModernBERT). The cross-encoder reranker (ms-marco-MiniLM-L-12-v2, 33MB) downloads on first use of the ripvec engine.

Usage

CLI

ripvec "error handling" .                              # Default ripvec engine
ripvec "form validation hooks" -n 5                    # Top 5 results
ripvec "database migration" --mode keyword             # BM25 only
ripvec "session encoding" --model modernbert --index   # ModernBERT with persistent index
ripvec --model modernbert --index --exclude-extensions=jsonl,md  # Skip noisy extensions
ripvec -i --model modernbert --index .                 # Interactive TUI (transformer engines only)

MCP server

{ "mcpServers": { "ripvec": { "command": "ripvec-mcp" } } }

Tools (8 retrieval + 9 LSP):

Engine selection is per-daemon via RIPVEC_MCP_ENGINE={ripvec,modernbert,bert}; default is ripvec. Tool schemas are stable across engines: index_status reports engine: "ripvec" and cache_location: "in-memory" under the ripvec engine, engine: "modernbert" and an on-disk path under transformer engines.

LSP server

ripvec-mcp --lsp   # serves LSP over stdio

Same binary, --lsp flag selects protocol.

Supported languages

19 tree-sitter grammars, 30 file extensions:

Unsupported file types get sliding-window plain-text chunking. The embedding model handles any language; tree-sitter just provides better chunk boundaries.

Acknowledgments

ripvec's static bi-encoder uses Model2Vec embeddings (potion-base-32M, potion-code-16M) from MinishLab, whose semble pipeline inspired the path-enriched BM25 and query-shape boosting design we ported to Rust and extended. Cross-encoder rerank uses ms-marco-MiniLM-L-12-v2. See CREDITS.md for the full ledger of what we used, what we ported, and what we built on top.

Limitations

• goToDefinition is best-effort: resolves by name matching and structural importance, not by type system analysis. Use dedicated LSPs (rust-analyzer, pyright, gopls) when you need exact resolution for overloaded symbols.
• Call graph is approximate: common names like new, run, render may resolve to the wrong definition. Cross-crate resolution limited to workspace members.
• Ripvec engine top-10 coherence on long-form prose: ModernBERT retains an edge on narrative corpora where a single source document is the right answer for every position in the top-10. Top-3 quality is competitive; coherent top-10 is not. If you're searching a legal archive or a book collection and need 10 contiguous hits from the same source, --model modernbert --index is the better tool.
• Cacheless cold start scales linearly: first-query indexing on the ripvec engine is O(files). At 92K files (Linux kernel) it's ~75s. Persistent transformer engines amortize this across runs but pay model-download and disk-cache costs.
• English-centric: both engines were trained primarily on English text. Queries and code comments in other languages will have lower recall.

Development

cargo fmt --check && cargo clippy --all-targets -- -D warnings && cargo test --workspace

See CLAUDE.md for detailed development conventions, architecture notes, and MCP tool namespace resolution.

Architecture

Cargo workspace with three crates:

Docs

• CREDITS.md: full attribution for models, libraries, and design inspiration
• Metal/MPS Architecture
• CUDA Architecture
• Development Learnings

License

Licensed under either of Apache-2.0 or MIT at your option.