// the find
wandb/wandbot
wandbot is a technical support bot for Weights & Biases' AI developer tools that can run in Discord, Slack, ChatGPT and Zendesk
wandbot is the support assistant Weights & Biases runs over its own docs, code, and example colabs, served as a FastAPI endpoint with Slack and Discord front ends. It is most useful as a worked example of a production docs-RAG setup (LLM query expansion, Chroma retrieval, Cohere reranking, then synthesis), and less so as a drop-in bot for another company's docs, since the ingestion and config layers are tied to W&B's sources.
src/wandbot/rag splits the pipeline into query enhancement, retrieval, and response synthesis, and the model calls sit behind LLM and EmbeddingModel classes in src/wandbot/models. Swapping a provider doesn't mean touching retrieval, which is the part homegrown support bots usually get tangled in. Evaluation works as a release gate: the README keeps a version-by-version correctness table, the eval runs five trials per sample before a prod release, and the release checklist requires an ingestion report for every knowledge-base update. Config is centralized in src/wandbot/configs (embedding model, dimensions, hosted vs local Chroma, ingestion sources), so the settings that change answer quality live in one folder instead of inline constants. Deployment is documented end to end, with local uvicorn plus run.sh and a Modal split into an API app and a bots app. The hourly cron that keeps the bots alive is a workaround, but at least it is written down.
The README contradicts itself on the synthesis model. The Features section says GPT-4o does response synthesis, the v1.3.1 note says Sonnet-3.7 does, and the v1.3.0 changelog reads 'GPT-4o for composing the response (was gpt-4o)'. Someone trying to reproduce a score needs the real answer, and the docs don't give it. The description promises ChatGPT and Zendesk, but the README mentions only Slack, Discord, and the HTTP API. The visible tree (truncated) has apps/discord and apps/slack and nothing else, so either those integrations live outside this repo or the description is ahead of the code. The eval numbers can't be checked from the repo. Scores link to W&B Weave evaluations in the wandbot workspace, the dataset is a Weave object ref, and the table mixes gpt-4-preview and gpt-4o judges across rows, so 91.2% and 82% aren't obviously comparable. Setup asks for a lot before anything answers. The env list covers OpenAI, Cohere, and W&B, but omits the Gemini key that query expansion calls, and it doesn't say which of the Slack, Discord, and Modal pieces are optional. Expect a few failed starts.