{"title":"Exercise: Building a Production RAG Pipeline","description":null,"content":"# Exercise 1: Answer a Question with RAG\n\nWire up the full RAG query path over a small corpus: embed the documents, embed the question, retrieve the top three chunks by cosine similarity, assemble them into a prompt, and have the model answer using only that context.\n\nThis is the query pipeline every RAG system runs on each request. The model never sees your whole knowledge base, only the few chunks the retriever selected, so the quality of the answer depends as much on retrieval as on the model. Instructing the model to answer strictly from the provided context is what keeps it from filling gaps with its own training data.\n\n### Requirements\n\n- Embed all documents with `text-embedding-3-small` and embed the question.\n- Retrieve the top three documents by cosine similarity.\n- Build a prompt that includes only those three chunks as context.\n- Instruct the model to answer using only the context, and to say it does not know if the context does not contain the answer.\n- Print the retrieved chunks and the final answer.\n\n\n**main.py**\n\n```python\nfrom openai import OpenAI\nimport math\n\nclient = OpenAI() # Uses OPENAI_API_KEY from environment\n\ndocuments = [\n \"Our refund policy allows returns within 30 days of purchase with a receipt.\",\n \"Standard shipping takes 5 to 7 business days within the continental US.\",\n \"Express shipping is delivered within 2 business days for an extra fee.\",\n \"Gift cards are non-refundable and cannot be exchanged for cash.\",\n \"To track an order, use the tracking link emailed when the order ships.\",\n \"Customer support is available Monday to Friday, 9am to 5pm Eastern.\",\n]\n\nquestion = \"How long do I have to return something I bought?\"\n\ndef cosine_similarity(a, b):\n dot = sum(x * y for x, y in zip(a, b))\n return dot / (math.sqrt(sum(x * x for x in a)) * math.sqrt(sum(y * y for y in b)))\n\n# TODO: embed the documents and the question with text-embedding-3-small\n\n# TODO: retrieve the indices of the top 3 documents by cosine similarity\n\n# TODO: build a context string from those 3 chunks and prompt the model\n# to answer using ONLY that context\n\n# TODO: print the retrieved chunks and the model's answer\n```\n\n\n### Expected Output\n\nThe retriever surfaces the refund chunk, and the model answers from it rather than from general knowledge.\n\n\n```plaintext\nRetrieved:\n - Our refund policy allows returns within 30 days of purchase with a receipt.\n - Gift cards are non-refundable and cannot be exchanged for cash.\n - To track an order, use the tracking link emailed when the order ships.\n\nAnswer: You have 30 days from the date of purchase to return an item, and you need the receipt.\n```\n\n\n### Solution\n\n\n**main.py**\n\n```python\nfrom openai import OpenAI\nimport math\n\nclient = OpenAI()\n\ndocuments = [\n \"Our refund policy allows returns within 30 days of purchase with a receipt.\",\n \"Standard shipping takes 5 to 7 business days within the continental US.\",\n \"Express shipping is delivered within 2 business days for an extra fee.\",\n \"Gift cards are non-refundable and cannot be exchanged for cash.\",\n \"To track an order, use the tracking link emailed when the order ships.\",\n \"Customer support is available Monday to Friday, 9am to 5pm Eastern.\",\n]\n\nquestion = \"How long do I have to return something I bought?\"\n\ndef cosine_similarity(a, b):\n dot = sum(x * y for x, y in zip(a, b))\n return dot / (math.sqrt(sum(x * x for x in a)) * math.sqrt(sum(y * y for y in b)))\n\n# Embed documents and question\ndoc_vectors = [\n item.embedding\n for item in client.embeddings.create(model=\"text-embedding-3-small\", input=documents).data\n]\nquestion_vector = client.embeddings.create(\n model=\"text-embedding-3-small\", input=question\n).data[0].embedding\n\n# Retrieve the top 3 documents by cosine similarity\nscored = sorted(\n range(len(documents)),\n key=lambda i: cosine_similarity(question_vector, doc_vectors[i]),\n reverse=True,\n)\ntop_indices = scored[:3]\nretrieved = [documents[i] for i in top_indices]\n\n# Assemble context and answer using only that context\ncontext = \"\\n\".join(f\"- {chunk}\" for chunk in retrieved)\nresponse = client.chat.completions.create(\n model=\"gpt-5.4-mini\",\n messages=[\n {\n \"role\": \"system\",\n \"content\": (\n \"Answer the question using only the provided context. \"\n \"If the context does not contain the answer, say you do not know.\"\n ),\n },\n {\"role\": \"user\", \"content\": f\"Context:\\n{context}\\n\\nQuestion: {question}\"},\n ],\n)\n\nprint(\"Retrieved:\")\nfor chunk in retrieved:\n print(f\" - {chunk}\")\nprint(f\"\\nAnswer: {response.choices[0].message.content}\")\n```\n\n\nThe pipeline has two halves that fail independently. Retrieval picks the evidence by embedding similarity, and generation turns that evidence into an answer. Pinning the model to the supplied context with the system prompt is what makes the system grounded: when the right chunk is retrieved the answer is correct, and when no chunk covers the question the model abstains instead of inventing a response.\n\n---\n\n# Exercise 2: Multi-Query Retrieval\n\nA single phrasing of a question may miss relevant chunks that use different words. Multi-query retrieval expands the question into several paraphrases, retrieves for each, and unions the results before answering. This widens recall so a vaguely-worded question still finds the right evidence.\n\n### Requirements\n\n- Ask the model to produce two paraphrases of the question.\n- Embed the documents once; for each query variant (original plus paraphrases) retrieve the top chunks.\n- Union the retrieved chunk indices, build context, and answer using only that context.\n\n\n**main.py**\n\n```python\nimport math\nfrom openai import OpenAI\n\nclient = OpenAI() # Uses OPENAI_API_KEY from environment\n\ndocuments = [\n \"Our refund policy allows returns within 30 days of purchase with a receipt.\",\n \"Standard shipping takes 5 to 7 business days within the continental US.\",\n \"Express shipping is delivered within 2 business days for an extra fee.\",\n \"To track an order, use the tracking link emailed when the order ships.\",\n \"Customer support is available Monday to Friday, 9am to 5pm Eastern.\",\n]\n\nquestion = \"When will my order get here?\"\n\ndef cosine_similarity(a, b):\n dot = sum(x * y for x, y in zip(a, b))\n return dot / (math.sqrt(sum(x * x for x in a)) * math.sqrt(sum(y * y for y in b)))\n\n# TODO: ask the model for 2 paraphrases of the question (query expansion)\n# TODO: embed the documents once; for each query variant retrieve the top 2 chunks\n# TODO: union the retrieved chunks, build context, and answer using only that context\n```\n\n\n### Expected Output\n\nThe paraphrases surface the shipping chunks the original wording might have missed, and the answer is grounded in them.\n\n\n```plaintext\nQueries: ['When will my order get here?', 'How long does delivery take?', 'What is the shipping time for my order?']\nAnswer: Standard shipping takes 5 to 7 business days, and express shipping arrives within 2 business days.\n```\n\n\n### Solution\n\n\n**main.py**\n\n```python\nimport math\nfrom openai import OpenAI\n\nclient = OpenAI() # Uses OPENAI_API_KEY from environment\n\ndocuments = [\n \"Our refund policy allows returns within 30 days of purchase with a receipt.\",\n \"Standard shipping takes 5 to 7 business days within the continental US.\",\n \"Express shipping is delivered within 2 business days for an extra fee.\",\n \"To track an order, use the tracking link emailed when the order ships.\",\n \"Customer support is available Monday to Friday, 9am to 5pm Eastern.\",\n]\n\nquestion = \"When will my order get here?\"\n\ndef cosine_similarity(a, b):\n dot = sum(x * y for x, y in zip(a, b))\n return dot / (math.sqrt(sum(x * x for x in a)) * math.sqrt(sum(y * y for y in b)))\n\ndoc_vectors = [item.embedding for item in client.embeddings.create(model=\"text-embedding-3-small\", input=documents).data]\n\n# Generate query paraphrases (query expansion)\nexpand = client.chat.completions.create(\n model=\"gpt-5.4-mini\",\n messages=[{\"role\": \"system\", \"content\": \"Rewrite the question as 2 alternative search queries, one per line.\"},\n {\"role\": \"user\", \"content\": question}],\n)\nqueries = [question] + [q.strip() for q in expand.choices[0].message.content.splitlines() if q.strip()]\n\ndef top2(query):\n qv = client.embeddings.create(model=\"text-embedding-3-small\", input=query).data[0].embedding\n ranked = sorted(range(len(documents)), key=lambda i: cosine_similarity(qv, doc_vectors[i]), reverse=True)\n return ranked[:2]\n\nretrieved = sorted({i for query in queries for i in top2(query)})\ncontext = \"\\n\".join(f\"- {documents[i]}\" for i in retrieved)\nanswer = client.chat.completions.create(\n model=\"gpt-5.4-mini\",\n messages=[{\"role\": \"system\", \"content\": \"Answer using only the context. If it is not covered, say you do not know.\"},\n {\"role\": \"user\", \"content\": f\"Context:\\n{context}\\n\\nQuestion: {question}\"}],\n)\nprint(\"Queries:\", queries)\nprint(\"Answer:\", answer.choices[0].message.content)\n```\n\n\nExpanding the query trades a cheap extra model call for better recall: each paraphrase probes the index from a different angle, and the union catches chunks any single phrasing would have ranked too low. Deduplicating the indices before building context keeps the prompt tight. This is one of the highest-impact upgrades to a basic RAG pipeline, because retrieval misses, not generation, are the most common cause of wrong answers.","pageType":"ai-engineering"}

Get Premium