Submission week 4

This is a submission for the Hacktoberfest Open-Source AI Challenge Week 4: Ghost in the Machine. Séance.dev is an application that allows users to summon an archived open-source repository and converse with its "ghost." By providing a GitHub URL, the app retrieves the repository's commits, issues, pull requests, and documentation, enabling users to ask questions based on the data left behind.

The answers are sourced directly from the repository, with every claim linking back to the relevant commit, issue, or comment. If the record lacks information on a user's question, the ghost will acknowledge it. Additionally, there's a "Resurrect" button. Users can select an open issue, and the application will draft a patch plan, including the likely cause, files involved, fix steps, risks, and a test idea - all linked to evidence from the repository's history.

The theme of the project is to take the "Ghost in the Machine" literally. Abandoned open-source projects are essentially ghosts: the code remains, but the people and reasoning behind it have vanished. Most of the reasoning is hidden in old issue threads that are seldom read. Séance.dev brings this information to the surface. It is essential to note that this is a grounded simulation built from public repository history; it is not the actual maintainers, and the ghost never invents quotes or opinions for them.

The demo showcases live interaction with pre-loaded repositories. It does not require an internet connection after the initial ingestion, emphasizing the preservation aspect of the project. The code is available on GitHub through the AVENGERSTHE2102 repository. The setup requires a few commands, as outlined in the README. The project was built using open-source components running locally, including an LLM (Qwen/Llama model), embeddings (nomic-embed-text), retrieval (SQLite + vector extension or Elasticsearch), a FastAPI backend with streaming answers over Server-Sent Events, and frontend development using React/JS with CSS effects.

The development process involved ingesting commits, issues, pull requests, and documentation from the GitHub API and caching them on disk. Commits, issues, and PRs are chunked accordingly, while documentation is split by headings. During question time, the top-k chunks are retrieved, handed to the model with strict grounding rules, and the model answers using only the provided context, citing every factual claim.

If the retrieval is weak, the model will state that the record is silent. The Resurrect pipeline follows a similar approach, fetching the issue thread, retrieving related commits and PRs, and producing a structured plan for addressing the issue.

One limitation identified during development is the quality of answers provided by smaller models and handling long issue threads and rate limits. However, these challenges were managed within the project's framework. The project highlights the importance of open innovation, as it operates without relying on a service that could vanish.

The code remains under the user's control, and the user can swap the "brain" by changing the model with a single configuration line, eliminating dependence on a vendor's pricing or deprecation schedule. The project is cost-effective, as ingesting hundreds of issues and asking hundreds of questions incur no per-token costs.