Tenant rights

A renter-rights helper has to be right: telling someone they have thirty days when they have three is not a small mistake. So this agent answers from a prebaked index of public HUD guidance, shows the source it is reading, and will not state an exact number that is not in the documents.

The whole stack runs on one NVIDIA key: Riva STT, a NIM LLM (meta/llama-3.3-70b-instruct), Riva TTS, and NIM embeddings. NVIDIA has no native LiveKit LLM plugin, so the NIM model is reached through the openai plugin pointed at the NIM endpoint. The openai client here is just NVIDIA's transport:

openai.LLM(model=NIM_LLM_MODEL, base_url=NIM_BASE_URL, api_key=api_key)

Grounding is not optional, so it does not ride on a tool the model might skip. On every turn, on_user_turn_completed embeds the question and retrieves the top passages:

result = retrieve(self._index, query_vec, k=3, floor=self._floor)
if result.covered:
    passages = "\n\n".join(
        f"[Source: {hit.source_label}]\n{hit.text}" for hit in result.hits
    )

When they cover the question, the agent adds them to the turn context as a system note that tells the model to lean on the passages, note when a rule varies by state, and never state an exact number that is not in them. The top source shows on a Card. The safety is in the guards: an empty or garbled turn clears the card and raises StopResponse so the model never answers with no grounding, and a failed embedding lookup tells the model in-context not to answer from general knowledge at all.

question = (new_message.text_content or "").strip()
if not question:
    _unmount_card(self._room)
    raise StopResponse()

The index is welded to the embedding model that built it. prewarm stamps the model id into the index and refuses to start on a mismatch, so build it once with uv run python build_index.py before dev, and rebuild if you change embeddings.

Build it from an empty folder in the full walkthrough, or talk to the finished agent at https://playground.mahimai.ca/demos/tenant-rights.