"Many use cases require GPT-3 to respond to user questions with insightful answers. For example, a customer support chatbot may need to provide answers to common questions. The GPT models have picked up a lot of general knowledge in training, but we often need to ingest and use a large library of more specific information.\n",
"\n",
"In this notebook we will demonstrate a method for enabling GPT-3 able to answer questions using a library of text as a reference, by using document embeddings and retrieval. We'll be using a dataset of Wikipedia articles about the 2020 Summer Olympic Games. Please see [this notebook](examples/fine-tuned_qa/olympics-1-collect-data.ipynb) to follow the data gathering process."
"Mariusz Przybylski is a professional footballer from Poland, and not much of a high jumper! Evidently GPT-3 needs some assistance here. \n",
"\n",
"The first issue to tackle is that the model is hallucinating an answer rather than telling us \"I don't know\". This is bad because it makes it hard to trust the answer that the model gives us! \n",
"prompt = \"\"\"Answer the question as truthfully as possible, and if you're unsure of the answer, say \"Sorry, I don't know\".\n",
"\n",
"Q: Who won the 2020 Summer Olympics men's high jump?\n",
"A:\"\"\"\n",
"\n",
"openai.Completion.create(\n",
" prompt=prompt,\n",
" temperature=0,\n",
" max_tokens=300,\n",
" top_p=1,\n",
" frequency_penalty=0,\n",
" presence_penalty=0,\n",
" model=COMPLETIONS_MODEL\n",
")[\"choices\"][0][\"text\"].strip(\" \\n\")"
]
},
{
"cell_type": "markdown",
"id": "1af18d66-d47a-496d-ae5f-4c5d53caa434",
"metadata": {},
"source": [
"To help the model answer the question, we provide extra contextual information in the prompt. When the total required context is short, we can include it in the prompt directly. For example we can use this information taken from Wikipedia. We update the initial prompt to tell the model to explicitly make use of the provided text."
"prompt = \"\"\"Answer the question as truthfully as possible using the provided text, and if the answer is not contained within the text below, say \"I don't know\"\n",
"Adding extra information into the prompt only works when the dataset of extra content that the model may need to know is small enough to fit in a single prompt. What do we do when we need the model to choose relevant contextual information from within a large body of information?\n",
"**In the remainder of this notebook, we will demonstrate a method for augmenting GPT-3 with a large body of additional contextual information by using document embeddings and retrieval.** This method answers queries in two steps: first it retrieves the information relevant to the query, then it writes an answer tailored to the question based on the retrieved information. The first step uses the [Embedding API](https://beta.openai.com/docs/guides/embeddings), the second step uses the [Completions API](https://beta.openai.com/docs/guides/completion/introduction).\n",
"* Preprocess the contextual information by splitting it into chunks and create an embedding vector for each chunk.\n",
"* On receiving a query, embed the query in the same vector space as the context chunks and find the context embeddings which are most similar to the query.\n",
"* Prepend the most relevant context embeddings to the query prompt.\n",
"* Submit the question along with the most relevant context to GPT, and receive an answer which makes use of the provided contextual information."
"We plan to use document embeddings to fetch the most relevant part of parts of our document library and insert them into the prompt that we provide to GPT-3. We therefore need to break up the document library into \"sections\" of context, which can be searched and retrieved separately. \n",
"Sections should be large enough to contain enough information to answer a question; but small enough to fit one or several into the GPT-3 prompt. We find that approximately a paragraph of text is usually a good length, but you should experiment for your particular use case. In this example, Wikipedia articles are already grouped into semantically related headers, so we will use these to define our sections. This preprocessing has already been done in [this notebook](examples/fine-tuned_qa/olympics-1-collect-data.ipynb), so we will load the results and use them."
"We preprocess the document sections by creating an embedding vector for each section. An embedding is a vector of numbers that helps us understand how semantically similar or different the texts are. The closer two embeddings are to each other, the more similar are their contents. See the [documentation on OpenAI embeddings](https://beta.api.openai.org/docs/guides/embeddings/) for more information.\n",
"\n",
"This indexing stage can be executed offline and only runs once to precompute the indexes for the dataset so that each piece of content can be retrieved later. Since this is a small example, we will store and search the embeddings locally. If you have a larger dataset, consider using a vector search engine like [Pinecone](https://www.pinecone.io/) or [Weaviate](https://github.com/semi-technologies/weaviate) to power the search.\n",
"\n",
"For the purposes of this tutorial we chose to use Curie embeddings, which are 4096-dimensional embeddings at a very good price and performance point. Since we will be using these embeddings for retrieval, we’ll use the \"search\" embeddings (see the [documentation](https://beta.api.openai.org/docs/guides/embeddings/))."
"So we have split our document library into sections, and encoded them by creating embedding vectors that represent each chunk. Next we will use these embeddings to answer our users' questions.\n",
"At the time of question-answering, to answer the user's query we compute the query embedding of the question and use it to find the most similar document sections. Since this is a small example, we store and search the embeddings locally. If you have a larger dataset, consider using a vector search engine like [Pinecone](https://www.pinecone.io/) or [Weaviate](https://github.com/semi-technologies/weaviate) to power the search."
"We can see that the most relevant document sections for each question are the summaries for the Men's and Women's high jump competitions - which is exactly what we would expect."
"# 3) Add the most relevant contexts to the query prompt\n",
"\n",
"Once we've calculated the most relevant pieces of context, we construct a prompt by simply prepending them to the supplied query. It is helpful to use a query separator to help the model distinguish between separate pieces of text."
" header = \"\"\"Answer the question as truthfully as possible using the provided context, and if the answer is not contained within the text below, say \"I don't know.\"\\n\\nContext:\\n\"\"\"\n",
" Answer the question as truthfully as possible using the provided context, and if the answer is not contained within the text below, say \"I don't know.\"\n",
"* The women's high jump event at the 2020 Summer Olympics took place on 5 and 7 August 2021 at the Japan National Stadium. Even though 32 athletes qualified through the qualification system for the Games, only 31 took part in the competition. This was the 22nd appearance of the event, having appeared at every Olympics since women's athletics was introduced in 1928.\n",
"* The men's high jump event at the 2020 Summer Olympics took place between 30 July and 1 August 2021 at the Olympic Stadium. 33 athletes from 24 nations competed; the total possible number depended on how many nations would use universality places to enter athletes in addition to the 32 qualifying through mark or ranking (no universality places were used in 2021). Italian athlete Gianmarco Tamberi along with Qatari athlete Mutaz Essa Barshim emerged as joint winners of the event following a tie between both of them as they cleared 2.37m. Both Tamberi and Barshim agreed to share the gold medal in a rare instance where the athletes of different nations had agreed to share the same medal in the history of Olympics. Barshim in particular was heard to ask a competition official \"Can we have two golds?\" in response to being offered a 'jump off'. Maksim Nedasekau of Belarus took bronze. The medals were the first ever in the men's high jump for Italy and Belarus, the first gold in the men's high jump for Italy and Qatar, and the third consecutive medal in the men's high jump for Qatar (all by Barshim). Barshim became only the second man to earn three medals in high jump, joining Patrik Sjöberg of Sweden (1984 to 1992).\n",
"* The men's triple jump event at the 2020 Summer Olympics took place between 3 and 5 August 2021 at the Japan National Stadium. Approximately 35 athletes were expected to compete; the exact number was dependent on how many nations use universality places to enter athletes in addition to the 32 qualifying through time or ranking (2 universality places were used in 2016). 32 athletes from 19 nations competed. Pedro Pichardo of Portugal won the gold medal, the nation's second victory in the men's triple jump (after Nelson Évora in 2008). China's Zhu Yaming took silver, while Hugues Fabrice Zango earned Burkina Faso's first Olympic medal in any event.\n",
"We have now obtained the document sections that are most relevant to the question. As a final step, let's put it all together to get an answer to the question.\n",
"\n",
"# 4) Answer the user's question based on the context.\n",
"Wow! By combining the Embeddings and Completions APIs, we have created a question-answering model which can answer questions using a large base of additional knowledge. It also understands when it doesn't know the answer! \n",
"\n",
"For this example we have used a dataset of Wikipedia articles, but that dataset could be replaced with books, articles, documentation, service manuals, or much much more. **We can't wait to see what you create with GPT-3!**\n",
"Our Q&A model is less prone to hallucinating answers, and has a better sense of what it does or doesn't know. This works when the information isn't contained in the context; when the question is nonsensical; or when the question is theoretically answerable but beyond GPT-3's powers!"
]
},
{
"cell_type": "code",
"execution_count": 31,
"id": "26a1a9ef-e1ee-4f80-a1b1-6164ccfa5bac",
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"Selected 3 document sections:\n",
"('France at the 2020 Summer Olympics', 'Taekwondo')\n",
"('2020 Summer Olympics medal table', 'Medal count')\n",
"('Taekwondo at the 2020 Summer Olympics – Qualification', 'Qualification summary')\n",
"\n",
"Q: What is the total number of medals won by France, multiplied by the number of Taekwondo medals given out to all countries?\n",
"A: I don't know.\n"
]
}
],
"source": [
"query = \"What is the total number of medals won by France, multiplied by the number of Taekwondo medals given out to all countries?\"\n",
