A documentation chatbot has two connected problems: turn a documentation collection into something searchable, then use the retrieved material to answer a question.
The upstream Chat LangChain checkout in my collection makes those stages visible. Its README describes document splitting, OpenAI embeddings, Weaviate, question rewriting, and streaming answers, with FastAPI and Next.js providing the application layers.
Ingestion determines what can be found
Documentation needs to be divided into units that remain useful when retrieved independently. A passage that is too small can lose its surrounding explanation. A large passage can include too much unrelated material.
The retained project uses embeddings and a vector database to organize those passages. That gives the question-answering stage a collection to search, but the original document links and boundaries remain important for understanding the result.
Conversation adds another step
A follow-up question may depend on an earlier turn. Question rewriting helps turn that conversational reference into something the retrieval step can use.
It is worth examining this separately from answer generation. A poor rewrite can retrieve the wrong context before the final model has a chance to respond.
Streaming affects the interface rather than the underlying truth of the answer. It can make the response feel more immediate while leaving retrieval quality and source support as separate concerns.
A reference architecture
This is an upstream implementation retained for learning. Personal customization and live operation are not established by the checkout.
Its value in my history is architectural: it shows how several individually understandable steps become a documentation assistant. It also provides a useful comparison with the smaller two-stage website Q&A experiment.