Chat with Docs

Overview

Retrieval-augmented question answering over your own documents. The template indexes three short documents into a vector collection, retrieves passages, and answers from those passages only, citing the source of each claim. It is the “Chat With Your Documents” template under Examples.

  1. Inputs - question, search (the text used for retrieval), and three document texts (doc_specs, doc_charging, doc_warranty), all nodetool.input.StringInput.
  2. Collection (vector.Collection) - A collection named aurora_docs that uses the Ollama nomic-embed-text embedding model. Run Ollama with that model available.
  3. Indexing (vector.IndexTextChunk, three nodes) - Each document goes into the collection with a document_id and a source metadata entry.
  4. Retrieval (vector.QueryText) - Queries the collection with the search text and returns the top 5 documents.
  5. Context - A Code node joins the passages, and a nodetool.text.Prompt combines them with the question.
  6. Answer (nodetool.agents.Agent) - Answers strictly from the passages, cites the passage heading after each claim, and replies “I couldn’t find that in the provided documents.” when the answer is absent. The template selects gpt-5-mini.
  7. Outputs - Answer and Retrieved Passages.

The search input is separate from the question. The shipped value, Aurora, appears in every document, so retrieval returns the whole corpus. Narrow it, for example to warranty, to retrieve one topic. Replace the three document inputs to use your own corpus.

Demo

Tags

rag, vectorstore, retrieval, llm, citations

Workflow Diagram

graph TD collection["Collection (aurora_docs)"] docs["StringInput x3 (documents)"] index["IndexTextChunk x3"] search["StringInput (search)"] query["QueryText"] join["Code (join passages)"] question["StringInput (question)"] prompt["Prompt"] agent["Agent"] answer["Output (Answer)"] passages["Output (Retrieved Passages)"] docs -->|text| index collection --> index collection --> query search -->|text| query query -->|documents| join join -->|CONTEXT| prompt question -->|QUESTION| prompt prompt --> agent --> answer join --> passages
Edit this page on GitHub