Awesome AI Agents › AI Assistants

zamalali/DeepGit

⭐ 912 Python repository created 2024-12-16

DeepGit is an agentic research workflow built on LangGraph that performs deep discovery across GitHub repositories. The user submits a plain language question, and an orchestrator agent then runs a relay of specialised steps. First an LLM expands the query into high-signal GitHub search tags. The same pass detects hardware constraints implied by wording such as GPU-poor, low-memory or mobile-only, and records them for later filtering. Candidate READMEs and documentation blocks are embedded with multi-dimensional ColBERT v2 token vectors, and MaxSim scoring surfaces fine-grained matches that single-vector embeddings miss. A lightweight MiniLM-L-6-v2 cross-encoder then reorders the top results for passage-level accuracy. A dependency filter inspects each repository requirements.txt or pyproject.toml and discards projects that cannot run on the declared hardware. Community and code signals are collected next, including stars, forks, issue cadence, commit history and quick code-quality metrics. Finally a multi-factor ranking merges the scores and returns a table with links, similarity percentages and badges marking repositories that run on CPU only. The stated goals are to surface capable but little-known open source tools, to give research-focused developers an intelligent discovery layer over GitHub, and to keep the whole workflow open source. A web interface accepts the raw query and displays the ranked results with the individual threshold scores, and a reduced DeepGit-lite variant runs without a GPU on a Hugging Face Space for quick trials. The project targets Python 3.11 or newer with pip 24 or newer, and supports several LLM providers selected through an LLM_PROVIDER environment variable, currently Groq with a DeepSeek R1 distilled Llama model as the default and MiniMax as an alternative, with automatic provider detection from available API keys when the variable is unset.

https://github.com/zamalali/DeepGit

deep-researchlanggraphgithub-searchcolbertrerankingretrievalpythonllmrepository-discovery