{"title":"Exercise: Deep Research Agents","description":null,"content":"# Exercise 1: Fan Out and Synthesize\n\nBuild the core of a deep research agent: break a broad question into focused sub-questions, answer each one independently, then synthesize the answers into a single coherent response.\n\nA deep research agent does not answer a big question in one shot. It decomposes the question into parts, gathers an answer for each part, and combines them. This fan-out-then-synthesize shape is what lets an agent cover more ground than a single prompt, because each sub-question gets the model's full attention before the results are merged.\n\n### Requirements\n\n- Ask the model to break a broad question into three focused sub-questions.\n- Answer each sub-question with its own model call.\n- Synthesize the sub-answers into one final response that addresses the original question.\n- Print the sub-questions, the sub-answers, and the final synthesis.\n\n\n**main.py**\n\n```python\nfrom openai import OpenAI\nimport json\n\nclient = OpenAI() # Uses OPENAI_API_KEY from environment\n\nquestion = \"What should a team consider before adopting a vector database?\"\n\ndef decompose(question: str) -> list:\n # TODO: ask the model for exactly 3 focused sub-questions as a JSON list of strings\n pass\n\ndef answer(subquestion: str) -> str:\n # TODO: answer a single sub-question concisely\n pass\n\n# TODO: decompose the question, answer each sub-question, then synthesize the\n# sub-answers into one final response. Print sub-questions, answers, and synthesis.\n```\n\n\n### Expected Output\n\nThe broad question becomes three focused parts, each answered, then combined.\n\n\n```plaintext\nSub-questions:\n 1. What query patterns and scale will the database need to support?\n 2. How do hosting, cost, and operational burden compare across options?\n 3. How well does it integrate with the existing stack and embedding models?\n\nSynthesis: Before adopting a vector database, evaluate your query patterns and scale, weigh hosting and operational cost, and confirm it integrates with your stack...\n```\n\n\n### Solution\n\n\n**main.py**\n\n```python\nfrom openai import OpenAI\nimport json\n\nclient = OpenAI()\n\nquestion = \"What should a team consider before adopting a vector database?\"\n\ndef decompose(question: str) -> list:\n response = client.chat.completions.create(\n model=\"gpt-5.4-mini\",\n messages=[\n {\n \"role\": \"system\",\n \"content\": 'Break the question into exactly 3 focused sub-questions. Reply as JSON: {\"subquestions\": [\"...\", \"...\", \"...\"]}.',\n },\n {\"role\": \"user\", \"content\": question},\n ],\n response_format={\"type\": \"json_object\"},\n )\n return json.loads(response.choices[0].message.content)[\"subquestions\"]\n\ndef answer(subquestion: str) -> str:\n response = client.chat.completions.create(\n model=\"gpt-5.4-mini\",\n messages=[\n {\"role\": \"system\", \"content\": \"Answer the question concisely in one or two sentences.\"},\n {\"role\": \"user\", \"content\": subquestion},\n ],\n )\n return response.choices[0].message.content.strip()\n\nsubquestions = decompose(question)\nprint(\"Sub-questions:\")\nfor i, sq in enumerate(subquestions, start=1):\n print(f\" {i}. {sq}\")\n\nanswers = [answer(sq) for sq in subquestions]\n\nfindings = \"\\n\".join(f\"Q: {sq}\\nA: {a}\" for sq, a in zip(subquestions, answers))\nsynthesis = client.chat.completions.create(\n model=\"gpt-5.4-mini\",\n messages=[\n {\"role\": \"system\", \"content\": \"Synthesize these findings into one coherent answer to the original question.\"},\n {\"role\": \"user\", \"content\": f\"Original question: {question}\\n\\nFindings:\\n{findings}\"},\n ],\n)\nprint(f\"\\nSynthesis: {synthesis.choices[0].message.content.strip()}\")\n```\n\n\nDecomposition is what makes the research deep: each sub-question is small enough for the model to answer well, and together they cover more of the question than a single prompt would. Answering them independently is also where real agents parallelize and attach retrieval. The synthesis step is not optional glue; it is where separate findings become one answer, and giving the synthesizer both the original question and all the findings is what keeps the result on target rather than a list of disconnected facts.\n\n---\n\n# Exercise 2: Close the Gap with a Follow-up\n\nGood research notices what it missed. After a first answer, have the model critique its own draft to find an important missing aspect, research that gap with a follow-up question, and fold the result into a stronger final answer. This self-critique loop is what makes iterative research deeper than a single pass.\n\n### Requirements\n\n- Produce an initial answer to the question.\n- Ask the model what important aspect the draft missed (one follow-up question).\n- Answer that follow-up, then synthesize the draft and the follow-up into a final answer.\n\n\n**main.py**\n\n```python\nfrom openai import OpenAI\nimport json\n\nclient = OpenAI() # Uses OPENAI_API_KEY from environment\n\nquestion = \"What should a team consider before adopting a vector database?\"\n\ndef answer(q):\n return client.chat.completions.create(model=\"gpt-5.4-mini\",\n messages=[{\"role\": \"user\", \"content\": q}]).choices[0].message.content.strip()\n\n# TODO: 1) get an initial answer.\n# TODO: 2) ask the model what important aspect is still MISSING (one follow-up question).\n# TODO: 3) answer that follow-up, then synthesize the initial answer + follow-up into a final answer.\n```\n\n\n### Expected Output\n\nThe critique surfaces a missing aspect (such as cost or operations), and the final answer incorporates it.\n\n\n```plaintext\nIdentified gap: What are the ongoing operational and cost considerations of running a vector database?\nFinal: Before adopting a vector database, weigh query patterns and scale, integration with your stack, and the ongoing operational burden and cost of running it...\n```\n\n\n### Solution\n\n\n**main.py**\n\n```python\nfrom openai import OpenAI\nimport json\n\nclient = OpenAI() # Uses OPENAI_API_KEY from environment\n\nquestion = \"What should a team consider before adopting a vector database?\"\n\ndef answer(q):\n return client.chat.completions.create(model=\"gpt-5.4-mini\",\n messages=[{\"role\": \"user\", \"content\": q}]).choices[0].message.content.strip()\n\ninitial = answer(question)\n\ngap = client.chat.completions.create(\n model=\"gpt-5.4-mini\",\n messages=[{\"role\": \"system\", \"content\": \"Given a question and a draft answer, reply with ONE follow-up question about an important aspect the draft missed.\"},\n {\"role\": \"user\", \"content\": f\"Question: {question}\\nDraft: {initial}\"}],\n).choices[0].message.content.strip()\nprint(\"Identified gap:\", gap)\nfollow_up_answer = answer(gap)\n\nfinal = client.chat.completions.create(\n model=\"gpt-5.4-mini\",\n messages=[{\"role\": \"system\", \"content\": \"Combine the draft and the follow-up into one complete answer to the original question.\"},\n {\"role\": \"user\", \"content\": f\"Question: {question}\\nDraft: {initial}\\nFollow-up Q: {gap}\\nFollow-up A: {follow_up_answer}\"}],\n)\nprint(\"Final:\", final.choices[0].message.content.strip())\n```\n\n\nLetting the model critique its own draft turns a one-shot answer into an iterative one: the gap question targets exactly what the first pass underweighted, and researching it adds genuine coverage rather than rephrasing. Synthesizing the draft with the follow-up keeps the final answer coherent instead of bolted-on. Repeating this critique-and-fill step is how deep-research agents converge on thorough answers; a budget or a round limit decides when to stop.","pageType":"ai-engineering"}

Get Premium