Bionic

Agentic RAG Introduction

Retrieval Augmented Generation pairs embeddings with grounded text generation. We detail each layer: chunking, storage, retrieval, and fusion into the final assistant response.

Best practices cover chunk sizes, metadata filters, and how to justify every generated sentence with citations in the UI.

Tooling at a Glance

  • list_datasets: Reads the prompt’s scope and returns every dataset the assistant is allowed to query so the model never guesses dataset IDs and always stays inside the tenant’s permissions.
  • list_dataset_files: Given a dataset_id, this function enumerates the actual files, sizes, and batch counts inside that dataset so the agent can pick the right sources before searching.
  • list_documents: This call enumerates ad-hoc documents shared in the current conversation (with real file_id values) so the agent can reference uploaded material alongside curated datasets.

How Agentic RAG Uses These Tools

  1. The assistant starts by calling list_datasets to understand which curated knowledge bases are attached to the user’s prompt.
  2. For any dataset that sounds relevant, it uses list_dataset_files to preview specific files and narrow retrieval to the ones that best match the question.
  3. If the user uploaded fresh context, list_documents provides the temporary file IDs so the agent can cross-check the conversation attachments.
  4. Only after these grounding steps does the agent run the retrieval/query pipeline, blend the cited facts into a draft answer, and return citations that map back to the dataset files or conversation documents surfaced through the tools above.

Example Prompt and Flow

User prompt: “I just uploaded the latest ‘Retail Banking FAQ’ PDF. Using that and any banking compliance datasets we already configured, draft a response explaining which mortgage programs are available to freelancers in Canada.”

  1. The model receives the prompt, recognizes it needs grounded information, and triggers list_datasets to see which banking datasets are linked to the assistant.
  2. It notices a dataset called “Banking Compliance” and calls list_dataset_files with that dataset_id to inspect which files (e.g., canada_mortgage_rules.md) might contain freelancer eligibility details.
  3. Because the user mentioned an uploaded PDF, the agent calls list_documents to obtain the temporary file_id for “Retail Banking FAQ” so it can be read later (via read_document).
  4. Armed with dataset file metadata and the ad-hoc document ID, the retrieval layer fetches the most relevant chunks, grounds the answer, and the language model drafts a response citing the specific dataset file and the uploaded PDF when referencing each mortgage rule.