Skip to content
Tutorials Center
My learning

Topics

Lesson 36 of 39 ยท Projects and Practice

Project: Research Assistant Agent

Build a research assistant agent end to end: plan sub-questions, search a document set, take notes, cite sources and write a report, with runnable Python stand-ins.

In this project you will build a research assistant: you give it a question, and it breaks the question into sub-questions, searches a small collection of documents, collects notes with sources, and writes a short report with citations. It combines ideas from earlier lessons: planning, retrieval (RAG), the orchestrator pattern (orchestrator and workers) and a final check for unsupported claims.

No real model or web search is available here. The planner, summarizer and writer are deterministic stand-ins, and the "web" is a tiny in-memory document set. Where a real system would call a model, the code marks the spot, so you can replace it later. The architecture is the real part.

The design

flowchart TD
  Q["Research question"] --> P["Planner: sub-questions"]
  P --> S["Searcher: find documents"]
  S --> N["Note taker: facts with sources"]
  N --> W["Writer: report with citations"]
  W --> C["Checker: every claim has a source"]
  C -->|"gaps"| P
  C -->|"ok"| R["Final report"]

A loop: if the checker finds gaps, the planner adds sub-questions.

Requirements:

  1. Split the question into sub-questions.
  2. Search the document set for each.
  3. Record notes as {fact, source} pairs.
  4. Write a report where each claim carries a source id.
  5. Refuse to include a claim without a source.
  6. Cap the work: at most three rounds and five notes per sub-question.
Python A tiny searchable corpus

Press Run (or Ctrl + Enter) to see the result here.

Program input ยท what you'd type while the program runs

Type the answers before you press Run, one per line. Each time the program reads a line (with input()), it gets the next line from this box.

The last query returns an empty list. That matters: the agent must cope with "nothing found".

Step 2: plan and take notes

The planner turns the question into sub-questions (a stand-in rule here). The note taker pulls the sentence from each document that best matches the sub-question and saves it with its source id.

Python Plan, search and take notes

Press Run (or Ctrl + Enter) to see the result here.

Program input ยท what you'd type while the program runs

Type the answers before you press Run, one per line. Each time the program reads a line (with input()), it gets the next line from this box.

Step 3: write with citations and check

The writer builds the report only from notes. The checker verifies that every sentence in the report ends with a known source id and that the cited note really contains that claim. A report line without a source is rejected.

Python Write a cited report and verify it

Press Run (or Ctrl + Enter) to see the result here.

Program input ยท what you'd type while the program runs

Type the answers before you press Run, one per line. Each time the program reads a line (with input()), it gets the next line from this box.

The checker catches the invented last sentence. With a real model the writer is the part that may invent things, which is exactly why you verify against the notes rather than trusting it.

Step 4: the full loop with a round limit

Put the pieces together: plan, gather notes, write, check. If sub-questions found nothing, report the gap honestly instead of filling it in.

Python The complete research loop

Press Run (or Ctrl + Enter) to see the result here.

Program input ยท what you'd type while the program runs

Type the answers before you press Run, one per line. Each time the program reads a line (with input()), it gets the next line from this box.

Extending the project

  • Replace the stand-in planner and writer with real model calls that return JSON, and keep the checker unchanged.
  • Swap search for the vector index from embeddings and vector search.
  • Run sub-questions in parallel with a thread pool.
  • Add a budget guard (see deploying agents).
  • Write evals: questions whose answers you know, plus unanswerable questions that should produce a decline (see evaluating agents).
  • Treat retrieved text as untrusted (see guardrails).

Check your understanding

Quick check ยท 4 questions

of 4 answered You scored / 4 ยท perfect!

  1. 1. Why must the report writer only use the collected notes?

    Grounding in notes makes verification possible.
  2. 2. What should the agent do when a sub-question finds no documents?

    An honest gap is better than a fabricated fact.
  3. 3. What does the checker verify?

    Lines without a matching note are flagged as unsupported.
  4. 4. Why cap the number of research rounds?

    Every loop needs a limit in code.

Summary

  • The assistant plans sub-questions, searches, records notes with sources, writes and checks.
  • Grounding the report in notes and verifying citations catches invented claims.
  • Missing information is reported, never filled in.
  • Caps on rounds and notes keep the run bounded.
  • Replace the stand-ins with real model calls and a real retriever to take it further.

Next: Project: Customer Support Agent