git séance: I asked the ghost of `request` why it died, and made it cite its sources

Описание и применение

git séance использует графовое представление истории репозитория request и локальную модель Gemma 4 для ответов на вопросы о принятии решений, автоматически цитируя исходные проблемы и коммиты.

Задача

Inherited or depended on an abandoned Node.js HTTP library and needed to understand historical design decisions before modifying it.

Результат

The system achieved 9/10 correct answers with 9/10 correct citations using a combined graph + vector retrieval approach, outperforming plain vector search (8/10).

Рабочий процесс

• 1. Ingest GitHub issues, PRs, comments via API and commits via a blobless clone.
• 2. Build node tables (Thread, Comment, Commit) and relationship tables (REPLIES, MENTIONS, REFERS, PART_OF).
• 3. Embed nodes with nomic‑embed‑text and store in LanceDB.
• 4. Query graph with Lance‑Graph Cypher to retrieve relevant threads, comments, and commits.
• 5. Feed retrieved evidence to Gemma 4 model to generate an answer.
• 6. Apply regex to verify each citation matches retrieved evidence, dropping invalid ones.
• 7. Return answer with valid citations.

Доказательства

git séance lets you ask that history a question. A local open-weight model (Gemma 4 through Ollama) answers as the project's ghost, but it may only speak from the history: every claim cites an issue ([#3142]) or a commit ([b66e4cd]), and the code removes any citation that isn't in what was actually retrieved.

Источник

The first project summoned is request, the Node.js HTTP library that went into maintenance mode in 2020: 3,477 issues and pull requests, 12,441 comments and 2,290 commits.

Источник

Gemma 4 E4B, about a minute per answer on a 6 GB GTX 1660 Ti.

Источник