repolix 0.1.0

Local-first codebase context engine — ask plain English questions about any Python codebase

codesight

Ask plain English questions about any Python codebase. Get answers with exact file and line citations. Runs entirely on your machine.

codesight index ./myrepo
codesight query "how does authentication work"

Searching... Generating answer... ── Answer ──────────────────────────────────────────────── authenticate_user() validates credentials by calling validate_token() [1] which checks expiry and signature. On success it creates a session via SessionService.create() [2]. ── Citations ───────────────────────────────────────────── [1] auth/validators.py:14-28 (validate_token) [2] auth/session.py:45-67 (SessionService.create)

[confidence: high · 5 chunks · index: ./myrepo/.codesight]

Your code never leaves your machine. No server. No accounts beyond an OpenAI API key.

---

Why codesight

Getting dropped into an unfamiliar codebase is painful. Documentation is outdated. Grep finds strings, not meaning. LLM chatbots hallucinate file names and function signatures because they have no access to your actual code.

codesight indexes your code locally using AST-based chunking — every retrieved chunk is a complete function or class, never an arbitrary line slice. It runs entirely on your machine.

---

How it works

1. AST chunking Tree-sitter parses each file into a syntax tree. codesight splits only at function and class boundaries. Every chunk is semantically complete. Methods are tracked with their parent class for disambiguation.

2. Hybrid search Queries run against OpenAI embeddings (vector search) and exact token matching (keyword search) simultaneously. Results are merged using Reciprocal Rank Fusion — a ranking algorithm that rewards consistency across search methods over dominance in just one.

3. Call graph expansion After initial retrieval, codesight inspects each retrieved chunk's call graph and fetches called functions that did not rank highly enough on their own. This surfaces implementation details that live one function call away from the entry point.

4. Metadata re-ranking Retrieved chunks are re-ranked using function names, file paths, docstrings, and call graph signals before being sent to the LLM.

5. Cited answers The top chunks go to gpt-5.4-mini with instructions to synthesize across all chunks and cite every claim. Citations map back to exact file paths and line numbers.

---

Quickstart

Requirements

• Python 3.11+
• OpenAI API key (get one here)

Node.js is not required for end users. The web UI is bundled inside the package and served directly by FastAPI.

Install from PyPI

pip install codesight

Set your API key:

export OPENAI_API_KEY=sk-your-key-here
# or add it to a .env file in your working directory

Install from source (development)

git clone https://github.com/TheAsianFish/codesight
cd codesight
python -m venv .venv
source .venv/bin/activate      # Windows: .venv\Scripts\activate
pip install -e ".[dev]"
cp .env.example .env
# Edit .env and add your OPENAI_API_KEY

CLI

# Index a repository (~$0.02 per 30k lines, one-time)
codesight index ./path/to/repo

# Ask a question
codesight query "how does authentication work"

# See raw retrieved chunks without an LLM call
codesight query "where is UserService defined" --no-llm

# Force re-index all files after a major refactor
codesight index ./path/to/repo --force

Web UI

# Start the server (the React UI is bundled — no npm needed)
uvicorn codesight.api:app --port 8000
# Open http://localhost:8000

For frontend development (hot reload via Vite):

# Requires Node.js 18+
cd frontend && npm install && cd ..
bash start.sh
# Backend: http://localhost:8000  Frontend: http://localhost:3000

---

Cost

Action	Cost
Index 30k line repo	~$0.02 (one-time)
Re-index after small change	~$0.001 (changed files only)
Each query	~$0.001

Incremental indexing means re-indexing after a small change costs almost nothing — only changed files are re-embedded.

---

Stack

Layer	Choice
AST parsing	Tree-sitter
Embeddings	text-embedding-3-small
Vector store	ChromaDB (local, no server needed)
LLM	gpt-5.4-mini
Backend	FastAPI
Frontend	React + TypeScript
CLI	Click

---

Output

Each query produces:

• A prose answer with inline citations [1], [2] etc.
• A citations section with exact file paths and line ranges. Citations marked [truncated] mean the function exceeded the 300-token chunk cap — the answer is based on a partial view of that function.
• A confidence label (high / medium / low) derived from how strongly the retrieved chunks matched the query across function names, file paths, docstrings, and call graph signals.

---

Limitations

• Python repos only. TypeScript support planned for V2.
• Best on repos up to ~30k lines.
• Deeply nested functions are included in their parent chunk.
• Large functions (>300 tokens) are truncated at the chunk cap. The [truncated] marker in citations flags when this occurs.
• Complex cross-file reasoning may require rephrasing the query.
• Architecture-level questions (layer structure, dependency graphs) require the V2 dependency graph feature to answer reliably.

---

Roadmap

V2 — TypeScript support, VS Code extension, dependency graph

V3 — GitHub webhook re-indexing, multi-repo, Slack bot

---

Contributing

Bug reports and pull requests are welcome. Please open an issue before submitting a large change so we can discuss the approach.

See .github/ISSUE_TEMPLATE/bug_report.md for the bug report format.

---

License

MIT © 2026 Patrick Chung