fva

FVA — FFF · Vector · AST

Hybrid codebase intelligence engine for AI coding agents. FVA combines FFF (fuzzy file search), vector embeddings, and AST chunking via Tree-sitter, plus call graphs, into a single MCP server — so agents search by path, content, meaning, and structure in one pass.

Why FVA

Most agent workflows chain grep → read_file → repeat. FVA replaces that loop with fused search:

• File discovery — frecency-ranked fuzzy paths (FFF)
• Semantic recall — embedding search over AST chunks
• Structural context — call graph neighbors and symbol bodies
• One MCP server — stdio transport, local-first, no telemetry

Install

Pre-built binaries are published on GitHub Releases for:

One-liner

Linux / macOS (Apple Silicon)

curl -fsSL https://raw.githubusercontent.com/Xeon-Dot/fva/main/scripts/install.sh | bash

Windows (PowerShell)

irm https://raw.githubusercontent.com/Xeon-Dot/fva/main/scripts/install.ps1 | iex

Installs to ~/.local/bin (Unix) or %LOCALAPPDATA%\Programs\fva\bin (Windows) and adds the directory to your user PATH.

Pin a version

FVA_VERSION=v0.2.0 curl -fsSL https://raw.githubusercontent.com/Xeon-Dot/fva/main/scripts/install.sh | bash

$env:FVA_VERSION = "v0.2.0"; irm https://raw.githubusercontent.com/Xeon-Dot/fva/main/scripts/install.ps1 | iex

Manual download

Download the archive for your platform from Releases, verify against SHA256SUMS.txt, and place fva (or fva.exe) on your PATH.

Build from source

Requires Rust 1.75+.

git clone https://github.com/Xeon-Dot/fva.git
cd FVA
cargo build --release

Binary: target/release/fva (.exe on Windows).

Intel Mac has no pre-built binary yet — use cargo install --path . or build from source.

Quick Start

# Verify install
fva --version

# Start MCP server (default mode)
fva --path /path/to/project

# Build full index (AST + vectors + call graph)
fva index --path .

# Check index status
fva status --path .

# CLI hybrid search
fva search "authentication handler" --path . --limit 5

On first run, FVA creates a .fva/ directory in the project root for indexes, frecency data, and vectors.

MCP Client Setup

Add FVA to your MCP client config. Use the installed binary path on your system.

Ready-to-copy examples for each agent tool live in examples/mcp-clients/. See manifest.json for install paths.

macOS / Linux (generic mcpServers format)

{
  "mcpServers": {
    "fva": {
      "command": "fva",
      "args": ["--path", "/path/to/your/project"],
      "env": { "RUST_LOG": "info" }
    }
  }
}

Windows

{
  "mcpServers": {
    "fva": {
      "command": "C:\\Users\\You\\AppData\\Local\\Programs\\fva\\bin\\fva.exe",
      "args": ["--path", "D:\\Dev\\YourProject"],
      "env": { "RUST_LOG": "info" }
    }
  }
}

If fva is on PATH, the short "command": "fva" form works on all platforms.

Recommended agent prompt

For codebase exploration, use FVA MCP tools:
- hybrid_search: default — combines file search + semantic + call graph
- semantic_search: natural language concept search
- get_smart_context: token-efficient context for a task
- get_symbol_info / get_chunks: full function/class bodies
- get_call_graph: callers and callees
Prefer hybrid_search over repeated grep+read cycles.

MCP Tools

Configuration

Copy config.example.toml to fva.toml (or .fva.toml) in the project root, and/or to ~/.config/fva/config.toml for global defaults. Project settings override global; CLI --config overrides both.

[embedding]
provider = "local"    # "local" (default) or "voyage"
model = "voyage-code-3"

[vector]
backend = "lancedb"    # "lancedb" (default, native LanceDB) | "lancedb-native" (alias)
db_path = "vectors"

[query]
fff_weight = 0.3
vector_weight = 0.5
graph_weight = 0.2
bm25_weight = 0.35
max_context_tokens = 8000

CLI flags override config: --path, --config, --log-level / RUST_LOG.

Voyage API (optional)

For higher-quality embeddings, switch provider and set your API key:

[embedding]
provider = "voyage"

export VOYAGE_API_KEY=your-key-here

Architecture

┌─────────────────────────────────────────────────────────────┐
│                    MCP Server (stdio)                        │
│  hybrid_search │ semantic_search │ grep │ find_files │ ... │
└─────────────┬───────────────────────┬───────────────────────┘
              │                       │
     ┌────────▼────────┐     ┌────────▼────────┐
     │   FFF Engine    │     │  AST Indexer    │
     │ frecency+fuzzy  │     │  Tree-sitter    │
     │  git-aware grep │     │  chunk store    │
     └────────┬────────┘     └────────┬────────┘
              │                ┌───────▼────────┐
              │                │  Vector Store  │
              │                │  (flat/LanceDB)│
              │                └───────┬────────┘
              │                ┌───────▼────────┐
              │                │  Call Graph    │
              │                │  (petgraph)    │
              └────────┬───────┴────────┬───────┘
                       │                │
                ┌──────▼────────────────▼──────┐
                │     Hybrid Query Engine      │
                │  FFF → Vector → Graph fusion  │
                └──────────────────────────────┘

Development Status

Module Structure

src/
├── main.rs            # CLI (serve, index, status, search)
├── engine.rs          # Central orchestrator
├── embedding/         # local-hash + voyage providers
├── vector/            # flat vector store (LanceDB optional)
├── graph/             # call graph (petgraph)
├── indexer/           # AST parsing + chunking + pipeline
├── query/             # hybrid search + smart context
├── fff/               # FFF integration
└── mcp/               # MCP tool handlers

Performance Targets

Security

• Sandboxed indexing — project root only
• No telemetry — all data stored locally in .fva/
• Embeddings are local by default; Voyage only when explicitly configured

License

MIT