Launch HN: Greptile (YC W24) - RAG on codebases that actually works

Hacker News Discussion ↗

A new tool called Greptile applies retrieval‑augmented generation (RAG) to entire codebases so developers can ask natural-language questions about projects and get context-aware answers, from how specific types are serialized to how different components interact. Commenters are intrigued by its potential for code comprehension, IDE integration, and use on large or multi-language repositories, but raise concerns about reliability at scale, GitHub permissions, privacy of embeddings, lack of metadata like issues/PRs, and the maturity of the product given outages and errors during launch. Several compare it to existing tools like Cursor, Cody, and Bloop, and request features such as self-hosting, better local workflows, and support for more public repos and documentation sources.

Product & Core Approach

  • Tool offers “RAG on codebases”: uses ASTs (via tree‑sitter), embeddings, and LLMs to answer questions about entire repos.
  • Focus is on code comprehension and internal-docs replacement/augmentation, not primarily code generation (though users do use it for code gen).
  • Currently indexes code only; plans to add commit messages, PRs, issues, comments, and better use of doc comments.

User Experiences & Use Cases

  • Positive reports: answered nuanced framework questions (e.g., Rails BigDecimal JSON encoding) matching what users learned manually.
  • Some see it as expanding LLM context window for complex projects; interest in using it on large Rails and multi-language repos.
  • Users want it to help with API error debugging if API source is available.

Integrations & Platforms

  • Has a VS Code extension; JetBrains plugin is on the roadmap.
  • Public demo: ~100 open‑source repos can be queried without login.
  • GitHub app required for private repos; some confusion about permissions and “act on your behalf” wording.

Reliability, Performance & UX Issues

  • Many reports of errors (“internal error while processing/locating sources”), failed or stuck processing (often at 99%), and AWS/DB outages, especially during HN traffic.
  • Bugs in popular-repo links, voting UI, branch/repo selection, and leftover branding (“Onboard”) noted.
  • Progress indicator and repo-selection UI considered misleading or fiddly.

Privacy, Security & Self‑Hosting

  • Claims to not store code after processing; currently stores embeddings of generated docstrings, with discussion that embeddings can leak information.
  • Some users would prefer explicit code storage if it improves speed, and/or fully local or self‑hosted versions.
  • Team sees self‑hosting and on‑prem as a future, but LLMs likely won’t be self‑hosted in near term.

Pricing, Limits & Adoption Friction

  • Free tier has repo size limits; users with large test datasets struggle to evaluate.
  • Suggestions: .greptileignore-style exclusions, data versioning tools, and more flexible trial (e.g., one large repo).
  • Concerns about automatic email use from GitHub; requests for opt‑in and separate email choice.

Positioning & Comparisons

  • Compared to Bloop, Adrenaline, Cursor, and Cody; pitched as full‑codebase comprehension rather than IDE replacement.
  • Multiple people flag that “RAG” jargon is unexplained or confusing and should be better messaged.