Reference

Product Introduction

1. Definition: Reference is a local semantic search engine and knowledge retrieval tool specifically designed for developers and AI agents. It operates as a privacy-first desktop application that indexes your local files and codebases.
2. Core Value Proposition: It exists to eliminate inefficient, context-less file searching (like grep loops) and provide AI agents with precise, cited access to a user's personal or project knowledge base. Its core value is delivering accurate code citations and semantic understanding without any data leaving the user's machine.

Main Features

1. Local Semantic Search with Citations: Reference performs vector search and keyword hybrid search on your local files. When you ask a natural language question like "how did I implement rate limiting," it returns the relevant code snippets or document sections with precise citations down to the function or line number, not just file names.
2. Live Indexing & Code-Aware Chunking: The index updates in real-time as you save files. It uses Tree-sitter, a robust parsing library, for language-aware code chunking. This means it intelligently splits code by functions, classes, or logical blocks, preserving structure for more accurate semantic retrieval compared to naive text splitting.
3. Built-in MCP (Model Context Protocol) Server: This is a key integration feature. Reference exposes endpoints like /search, /explain, /find_similar, and /check_doc_drift via MCP. This allows AI agents like Claude Code to directly query your local knowledge base, retrieving cited context to inform their actions without wasting tokens on manual file discovery.
4. Privacy-First, Offline Architecture: All processing—including embedding generation for semantic search—occurs locally. There are no cloud uploads, ensuring complete data privacy and security for proprietary code and sensitive documents.

Problems Solved

1. Pain Point: The inefficiency and inaccuracy of traditional search (grep, find) and the token-wasting behavior of AI assistants performing blind, iterative file searches. It also addresses the context loss from providing AI with incomplete code snippets.
2. Target Audience: Software Engineers and Developers working on large codebases; DevOps Engineers managing scripts and configs; Technical Leads onboarding team members; AI Agent Developers building tools that need contextual grounding; Security-Conscious Professionals in regulated industries.
3. Use Cases: Quickly recalling a specific implementation detail during code review; onboarding to a new codebase by asking "how does authentication work here?"; providing an AI coding assistant with exact, cited context to generate a patch or fix; checking for documentation drift between comments and actual code; finding all similar utility functions across a monorepo.

Unique Advantages

1. Differentiation: Unlike cloud-based code search tools (e.g., Sourcegraph Cloud) or generic desktop search, Reference is 100% local and privacy-focused. Unlike plain grep, it understands semantic meaning. Unlike feeding entire files to an AI, it provides pinpoint citations, reducing noise and cost.
2. Key Innovation: The integration of a built-in MCP server transforms it from a passive search tool into an active context provider for AI agents. This "pluggable memory" architecture allows AI workflows to be grounded in a user's exact, up-to-date codebase dynamically, which is a significant leap over static context windows or manual copy-pasting.

Frequently Asked Questions (FAQ)

1. How does Reference ensure my code privacy? Reference is a fully offline application; all indexing, embedding, and search processing occurs on your local machine. No code, file contents, or metadata is ever sent to an external server or cloud service.
2. What is the MCP server and how do AI agents use it? The Model Context Protocol (MCP) is a standard for tools to provide context to AI models. Reference's MCP server allows compatible AI agents (like Claude Desktop with MCP) to programmatically perform searches, explain code, and find similar code within your local files, receiving cited results directly within the agent's workflow.
3. What programming languages and file types does Reference support? Reference uses Tree-sitter for parsing, which provides strong support for most major languages (Python, JavaScript, Go, Rust, Java, etc.). For non-code files, it uses intelligent plain-text chunking. The live index adapts as you add files in supported formats.
4. How is semantic search in Reference different from just using 'grep'? grep only matches exact keywords. Reference's semantic search uses machine learning models to understand the meaning of your query and find conceptually related content, even if the terminology doesn't match exactly. It combines this with keyword search for robust results.
5. Can Reference be used for team or enterprise codebases? While primarily a desktop tool, its architecture makes it suitable for individuals working within large enterprise codebases. Each developer runs their own instance, ensuring security and allowing personalized indexing of relevant project subsets without central infrastructure.