Teams AI Library Vector Data Source RAG Grounding Prompt

You are a senior platform engineer who builds Microsoft Teams automation and grounded conversational assistants using the Teams AI Library.

I will provide:
- The knowledge corpus (runbooks, postmortems, on-call docs) and where it lives today
- The embedding model and vector store I intend to use (or want recommended)
- The bot's prompt template configuration and the planner type currently registered

Your job:

1. **Implement the DataSource interface** — write a class that satisfies the Teams AI Library DataSource contract (renderData), returning ranked, token-budgeted text plus source identifiers.
2. **Design the ingestion pipeline** — chunk the corpus with overlap, generate embeddings, and store vectors with metadata (doc id, section, lastUpdated) so retrieval can filter on freshness.
3. **Register the source in the prompt** — show how to reference the data source in the prompt's config.json/skprompt so the Action Planner augments the system context before planning.
4. **Enforce a token budget** — cap retrieved context to fit the model window alongside conversation history, and explain the trade-off between recall and prompt size.
5. **Add citations and grounding guardrails** — require the model to cite retrieved chunk ids and to answer "not found in the runbooks" rather than hallucinate when retrieval is empty.
6. **Plan re-indexing and staleness handling** — define how updated docs trigger re-embedding and how stale chunks are evicted.

Output as: the DataSource implementation code, the ingestion pipeline outline, the prompt config snippet wiring it in, and a short grounding-quality test plan.

Retrieval grounding reduces but does not eliminate hallucination — keep the "answer only from retrieved context" instruction explicit and test empty-retrieval behavior.