"[![Open In Collab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/langchain-ai/langchain/blob/master/docs/extras/use_cases/question_answering/qa.ipynb)\n",
"\n",
"## Use case\n",
"Suppose you have some text documents (PDF, blog, Notion pages, etc.) and want to ask questions related to the contents of those documents. LLMs, given their proficiency in understanding text, are a great tool for this.\n",
"\n",
"In this walkthrough we'll go over how to build a question-answering over documents application using LLMs. Two very related use cases which we cover elsewhere are:\n",
"- [QA over structured data](/docs/use_cases/sql) (e.g., SQL)\n",
"The pipeline for converting raw unstructured data into a QA chain looks like this:\n",
"1. `Loading`: First we need to load our data. Unstructured data can be loaded from many sources. Use the [LangChain integration hub](https://integrations.langchain.com/) to browse the full set of loaders.\n",
"Each loader returns data as a LangChain [`Document`](/docs/components/schema/document).\n",
"2. `Splitting`: [Text splitters](/docs/modules/data_connection/document_transformers/) break `Documents` into splits of specified size\n",
"3. `Storage`: Storage (e.g., often a [vectorstore](/docs/modules/data_connection/vectorstores/)) will house [and often embed](https://www.pinecone.io/learn/vector-embeddings/) the splits\n",
"4. `Retrieval`: The app retrieves splits from storage (e.g., often [with similar embeddings](https://www.pinecone.io/learn/k-nearest-neighbor/) to the input question)\n",
"5. `Generation`: An [LLM](/docs/modules/model_io/models/llms/) produces an answer using a prompt that includes the question and the retrieved data\n",
"6. `Conversation` (Extension): Hold a multi-turn conversation by adding [Memory](/docs/modules/memory/) to your QA chain.\n",
"\n",
"![flow.jpeg](/img/qa_flow.jpeg)\n",
"\n",
"## Quickstart\n",
"\n",
"To give you a sneak preview, the above pipeline can be all be wrapped in a single object: `VectorstoreIndexCreator`. Suppose we want a QA app over this [blog post](https://lilianweng.github.io/posts/2023-06-23-agent/). We can create this in a few lines of code. First set environment variables and install packages:"
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "e14b744b",
"metadata": {},
"outputs": [],
"source": [
"pip install openai chromadb\n",
"\n",
"# Set env var OPENAI_API_KEY or load from a .env file\n",
"' Task decomposition is a technique used to break down complex tasks into smaller and simpler steps. It can be done using LLM with simple prompting, task-specific instructions, or with human inputs. Tree of Thoughts (Yao et al. 2023) is an extension of Chain of Thought (Wei et al. 2022) which explores multiple reasoning possibilities at each step.'"
]
},
"execution_count": 3,
"metadata": {},
"output_type": "execute_result"
}
],
"source": [
"index.query(\"What is Task Decomposition?\")"
]
},
{
"cell_type": "markdown",
"id": "8224aad6",
"metadata": {},
"source": [
"Ok, but what's going on under the hood, and how could we customize this for our specific use case? For that, let's take a look at how we can construct this pipeline piece by piece."
]
},
{
"cell_type": "markdown",
"id": "ba5daed6",
"metadata": {},
"source": [
"## Step 1. Load\n",
"\n",
"Specify a `DocumentLoader` to load in your unstructured data as `Documents`. A `Document` is a piece of text (the `page_content`) and associated metadata."
"To be able to look up our document splits, we first need to store them where we can later look them up.\n",
"The most common way to do this is to embed the contents of each document then store the embedding and document in a vector store, with the embedding being used to index the document."
"Vectorstores are commonly used for retrieval, but they are not the only option. For example, SVMs (see thread [here](https://twitter.com/karpathy/status/1647025230546886658?s=20)) can also be used.\n",
"\n",
"LangChain [has many retrievers](/docs/modules/data_connection/retrievers/) including, but not limited to, vectorstores. All retrievers implement a common method `get_relevant_documents()` (and its asynchronous variant `aget_relevant_documents()`)."
"Some common ways to improve on vector similarity search include:\n",
"- `MultiQueryRetriever` [generates variants of the input question](/docs/modules/data_connection/retrievers/MultiQueryRetriever) to improve retrieval.\n",
"- `Max marginal relevance` selects for [relevance and diversity](https://www.cs.cmu.edu/~jgc/publication/The_Use_MMR_Diversity_Based_LTMIR_1998.pdf) among the retrieved documents.\n",
"- Documents can be filtered during retrieval using [`metadata` filters](/docs/use_cases/question_answering/how_to/document-context-aware-QA)."
]
},
{
"cell_type": "code",
"execution_count": 9,
"id": "c690f01a",
"metadata": {},
"outputs": [
{
"name": "stderr",
"output_type": "stream",
"text": [
"INFO:langchain.retrievers.multi_query:Generated queries: ['1. How can Task Decomposition be approached?', '2. What are the different methods for Task Decomposition?', '3. What are the various approaches to decomposing tasks?']\n"
"Distill the retrieved documents into an answer using an LLM/Chat model (e.g., `gpt-3.5-turbo`) with `RetrievalQA` chain.\n"
]
},
{
"cell_type": "code",
"execution_count": 10,
"id": "99fa1aec",
"metadata": {},
"outputs": [
{
"data": {
"text/plain": [
"{'query': 'What are the approaches to Task Decomposition?',\n",
" 'result': 'There are three approaches to task decomposition:\\n\\n1. Using Language Model with simple prompting: This approach involves using a Language Model (LLM) with simple prompts like \"Steps for XYZ\" or \"What are the subgoals for achieving XYZ?\" to guide the task decomposition process.\\n\\n2. Using task-specific instructions: In this approach, task-specific instructions are provided to guide the task decomposition. For example, for the task of writing a novel, an instruction like \"Write a story outline\" can be given to help decompose the task into smaller subtasks.\\n\\n3. Human inputs: Task decomposition can also be done with the help of human inputs. This involves getting input and guidance from humans to break down a complex task into smaller, more manageable subtasks.'}"
"Note, you can pass in an `LLM` or a `ChatModel` (like we did here) to the `RetrievalQA` chain.\n",
"\n",
"### Go deeper\n",
"\n",
"#### Choosing LLMs\n",
"- Browse the > 55 LLM and chat model integrations [here](https://integrations.langchain.com/).\n",
"- See further documentation on LLMs and chat models [here](/docs/modules/model_io/models/).\n",
"- Use local LLMS: The popularity of [PrivateGPT](https://github.com/imartinez/privateGPT) and [GPT4All](https://github.com/nomic-ai/gpt4all) underscore the importance of running LLMs locally.\n",
"Using `GPT4All` is as simple as [downloading the binary]((/docs/integrations/llms/gpt4all)) and then:"
]
},
{
"cell_type": "code",
"execution_count": 11,
"id": "02d6c9dc",
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"Found model file at /Users/rlm/Desktop/Code/gpt4all/models/nous-hermes-13b.ggmlv3.q4_0.bin\n"
]
},
{
"name": "stderr",
"output_type": "stream",
"text": [
"objc[61331]: Class GGMLMetalClass is implemented in both /Users/rlm/miniforge3/envs/llama/lib/python3.9/site-packages/gpt4all/llmodel_DO_NOT_MODIFY/build/libreplit-mainline-metal.dylib (0x2e3384208) and /Users/rlm/miniforge3/envs/llama/lib/python3.9/site-packages/gpt4all/llmodel_DO_NOT_MODIFY/build/libllamamodel-mainline-metal.dylib (0x2e37b0208). One of the two will be used. Which one is undefined.\n",
"llama.cpp: using Metal\n",
"llama.cpp: loading model from /Users/rlm/Desktop/Code/gpt4all/models/nous-hermes-13b.ggmlv3.q4_0.bin\n",
"llama_model_load_internal: format = ggjt v3 (latest)\n",
"The full set of retrieved documents used for answer distillation can be returned using `return_source_documents=True`."
]
},
{
"cell_type": "code",
"execution_count": 14,
"id": "60004293",
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"4\n"
]
},
{
"data": {
"text/plain": [
"Document(page_content='Task decomposition can be done (1) by LLM with simple prompting like \"Steps for XYZ.\\\\n1.\", \"What are the subgoals for achieving XYZ?\", (2) by using task-specific instructions; e.g. \"Write a story outline.\" for writing a novel, or (3) with human inputs.', metadata={'description': 'Building agents with LLM (large language model) as its core controller is a cool concept. Several proof-of-concepts demos, such as AutoGPT, GPT-Engineer and BabyAGI, serve as inspiring examples. The potentiality of LLM extends beyond generating well-written copies, stories, essays and programs; it can be framed as a powerful general problem solver.\\nAgent System Overview In a LLM-powered autonomous agent system, LLM functions as the agent’s brain, complemented by several key components:', 'language': 'en', 'source': 'https://lilianweng.github.io/posts/2023-06-23-agent/', 'title': \"LLM Powered Autonomous Agents | Lil'Log\"})"
"Answer citations can be returned using `RetrievalQAWithSourcesChain`."
]
},
{
"cell_type": "code",
"execution_count": 16,
"id": "948f6d19",
"metadata": {},
"outputs": [
{
"data": {
"text/plain": [
"{'question': 'What are the approaches to Task Decomposition?',\n",
" 'answer': 'The approaches to Task Decomposition include:\\n1. Using LLM with simple prompting, such as providing steps or subgoals for achieving a task.\\n2. Using task-specific instructions, such as providing a specific instruction like \"Write a story outline\" for writing a novel.\\n3. Using human inputs to decompose the task.\\nAnother approach is the Tree of Thoughts, which extends the Chain of Thought (CoT) technique by exploring multiple reasoning possibilities at each step and generating multiple thoughts per step, creating a tree structure. The search process can be BFS or DFS, and each state can be evaluated by a classifier or majority vote.\\nSources: https://lilianweng.github.io/posts/2023-06-23-agent/',\n",
"Retrieved documents can be fed to an LLM for answer distillation in a few different ways.\n",
"\n",
"`stuff`, `refine`, `map-reduce`, and `map-rerank` chains for passing documents to an LLM prompt are well summarized [here](/docs/modules/chains/document/).\n",
" \n",
"`stuff` is commonly used because it simply \"stuffs\" all retrieved documents into the prompt.\n",
"\n",
"The [load_qa_chain](/docs/use_cases/question_answering/how_to/question_answering.html) is an easy way to pass documents to an LLM using these various approaches (e.g., see `chain_type`)."
]
},
{
"cell_type": "code",
"execution_count": 17,
"id": "29aa139f",
"metadata": {},
"outputs": [
{
"data": {
"text/plain": [
"{'output_text': 'The approaches to task decomposition mentioned in the provided context are:\\n\\n1. Chain of thought (CoT): This approach involves instructing the language model to \"think step by step\" and decompose complex tasks into smaller and simpler steps. It enhances model performance on complex tasks by utilizing more test-time computation.\\n\\n2. Tree of Thoughts: This approach extends CoT by exploring multiple reasoning possibilities at each step. It decomposes the problem into multiple thought steps and generates multiple thoughts per step, creating a tree structure. The search process can be BFS or DFS, and each state is evaluated by a classifier or majority vote.\\n\\n3. LLM with simple prompting: This approach involves using a language model with simple prompts like \"Steps for XYZ\" or \"What are the subgoals for achieving XYZ?\" to perform task decomposition.\\n\\n4. Task-specific instructions: This approach involves providing task-specific instructions to guide the language model in decomposing the task. For example, providing the instruction \"Write a story outline\" for the task of writing a novel.\\n\\n5. Human inputs: Task decomposition can also be done with human inputs, where humans provide guidance and input to break down the task into smaller subtasks.'}"