Clera
A documentation assistant that answers from official docs and shows you where the answer came from.
The problem
Ask a general-purpose chatbot a language question and you often get an answer that blends three versions of the API, one deprecated pattern, and occasionally a method that has never existed. Learners need answers tied to the official documentation with the source attached, and not only for whichever two languages someone decided to support.
What I built
Clera started as a docs-specialized assistant over Python and Java. A Django pipeline handles query normalization, intent detection, retrieval from pgvector, and grounded generation through Qwen 2.5-7B-Instruct. In April 2026 I made it general: an admin points the crawler at a documentation site, the pipeline strips the content down, embeds it, and stores the vectors. Answers carry their sources. When retrieval finds nothing useful, the model still replies, but says plainly that it is answering without a source.
- v1: Python and Java knowledge bases built from official documentation
- v2, April 2026: any language supported once an admin crawls its docs
- Crawl, clean, embed, and store pipeline writing into PostgreSQL with pgvector
- Intent detection, cosine-similarity retrieval, then grounded generation
- Qwen 2.5-7B-Instruct through Hugging Face, with citations in the chat UI
- React and Vite frontend with syntax highlighting and source browsing
How it turned out
It went from a two-language assistant to something an admin extends by pointing it at a website. The answers stay tied to documentation instead of the open web, which was the whole reason to build it.