|
1 | | -# chromaHackpack |
| 1 | +# Chroma Hackpack |
| 2 | + |
| 3 | +# Overview |
| 4 | + |
| 5 | +Chroma is a vector database. In this hackpack, we'll use it to implement |
| 6 | +retrieval augmented generation (RAG) – a technique for enhancing large language |
| 7 | +model (LLM) informational capabilities. Specifically, we'll build a chat bot to |
| 8 | +answer logistic questions about TreeHacks 2024. |
| 9 | + |
| 10 | +## Motivation |
| 11 | + |
| 12 | +LLMs like ChatGPT are capable |
| 13 | +of solving sophisticated tasks. However, their knowledge of current events and |
| 14 | +new information is often limited by training cut-off dates. Moreover, LLMs can |
| 15 | +exhibit hallucinatory behavior. In other words, LLMs have strong reasoning |
| 16 | +abilities but they often need the appropriate facts to reason with. |
| 17 | + |
| 18 | +Rerieval augmentented generation is a technique of **retrieving** information and |
| 19 | +then providing it to the LLM to **augment** the content it next **generates**. |
| 20 | +This helps mitigate hallucination and supplements the LLM's existing knowledge |
| 21 | +with facts of the developer's choice. |
| 22 | + |
| 23 | +In the case of our application, we'll use RAG to provide our chat bot up-to-date |
| 24 | +information regarding TreeHacks 2024. |
| 25 | + |
| 26 | +## How does RAG work? |
| 27 | + |
| 28 | +Typically, LLMs directly respond to a user's query. Retrievel augmented |
| 29 | +generation modifies the query by including relevant facts to the query. |
| 30 | + |
| 31 | +First, we select several documents containing information relevant to TreeHacks |
| 32 | +logistics. We then calculate embeddings for the document contents. Embeddings |
| 33 | +are vectors that represent the semantics of a given string. If two vector |
| 34 | +embeddings are similar, then we know the semantics of the two respective strings |
| 35 | +are also similar. These embeddings are all loaded into our vector database, |
| 36 | +Chroma. |
| 37 | + |
| 38 | +Once our vector database is populated, we can begin querying it. When the user |
| 39 | +prompts our chat bot, the following occurs: |
| 40 | + |
| 41 | +(1) Take in user input. |
| 42 | +(2) Pass the input's embedding into a vector database. Retrieve the `k` most |
| 43 | +similar vectors and their associated strings. Each of these strings represent |
| 44 | +the information that is most relevant to the user's query. |
| 45 | +(3) Pass the user's original input along with the information from the vector |
| 46 | +database into the LLM. |
| 47 | +(4) Return the LLM's output. |
| 48 | + |
| 49 | + |
| 50 | + |
| 51 | +This framework is simple, but powerful. There are several ways to introduce |
| 52 | +additional sophistication into RAG, but for the purpose of this hackpack we'll |
| 53 | +focus on the basics. |
| 54 | + |
| 55 | +# Project Walkthrough |
| 56 | + |
| 57 | +## Step 0: Installing Dependencies |
| 58 | + |
| 59 | +Ensure you have Python 3+ installed on your computer. |
| 60 | + |
| 61 | +Download this repository and run `pip install -r requirements.txt`. |
| 62 | + |
| 63 | +## Step 1: Setting up Chroma |
| 64 | + |
| 65 | +Before we can process user queries, we must populate a Chroma vector database |
| 66 | +with embeddings. |
| 67 | + |
| 68 | +### Pre-processing |
| 69 | + |
| 70 | +We will use the `langchain` library to load and pre-process our data. |
| 71 | + |
| 72 | +```python |
| 73 | +from langchain_community.document_loaders import DirectoryLoader |
| 74 | +from langchain.text_splitter import RecursiveCharacterTextSplitter |
| 75 | +``` |
| 76 | + |
| 77 | +First, we'll use the `DirectoryLoader` to load all the files from our |
| 78 | +`documents` folder. Then, we'll use the `RecursiveCharacterTextSplitter` to |
| 79 | +break each document down into a series of strings. Each string will have its own |
| 80 | +embedding and thus can be independently queried. |
| 81 | + |
| 82 | +```python |
| 83 | +loader = DirectoryLoader('./documents') |
| 84 | +documents = loader.load() |
| 85 | + |
| 86 | +text_splitter = RecursiveCharacterTextSplitter(chunk_size=1000, chunk_overlap=0) |
| 87 | +texts = text_splitter.split_documents(documents) |
| 88 | +``` |
| 89 | + |
| 90 | +You'll notice that the `documents` folder comes pre-populated with TreeHacks |
| 91 | +related documents. You may replace these documents if you'd like your RAG LLM to be |
| 92 | +fed a different set of information. |
| 93 | + |
| 94 | +### Loading into Chroma |
| 95 | + |
| 96 | +Setting up our Chroma database is very easy. |
| 97 | + |
| 98 | +```python |
| 99 | +from langchain_community.vectorstores import Chroma |
| 100 | +from langchain_openai import OpenAIEmbeddings |
| 101 | +``` |
| 102 | + |
| 103 | +We will use `OpenAIEmbeddings` to embed our texts. However, we don't need to |
| 104 | +manually do this – Chroma will handle it. We simply declare our Chroma database |
| 105 | +with the texts and the embedding function. |
| 106 | + |
| 107 | +```python |
| 108 | +embeddings = OpenAIEmbeddings() |
| 109 | +vectordb = Chroma.from_documents(documents=texts, embedding=embeddings, persist_directory='db') |
| 110 | +``` |
| 111 | + |
| 112 | +This will automatically produce a Chroma vector database containing all the text |
| 113 | +documents and their vector embeddings. If you'd like to use a [different |
| 114 | +embedding function](https://python.langchain.com/docs/integrations/text_embedding), you can easily replace it. |
| 115 | + |
| 116 | +Before running this code, you will need to set up your API key. Use this |
| 117 | +[tutorial](https://platform.openai.com/docs/quickstart/step-2-set-up-your-api-key) |
| 118 | +provided by OpenAI. |
| 119 | + |
| 120 | +Moreover, notice that we pass a value for `persist_directory`. This tells Chroma |
| 121 | +to locally save the vector database to the folder `db`. By doing so, we can |
| 122 | +simply load in the vector database next time we run our program. This allows us |
| 123 | +to avoid recomputing the embeddings for all the documents. |
| 124 | + |
| 125 | +## Step 2: Running Queries |
| 126 | + |
| 127 | +Now that we've configured our Chroma database, we'd like to query it for the |
| 128 | +purpose of RAG. `langchain` gives us a pre-packaged object to do this. |
| 129 | + |
| 130 | +```python |
| 131 | +from langchain_openai import OpenAI |
| 132 | +from langchain.chains import VectorDBQA |
| 133 | +``` |
| 134 | + |
| 135 | +The `VectorDBQA` automatically coordinates interactions between our LLM and |
| 136 | +Chroma vector database. We can declare it easily. |
| 137 | + |
| 138 | +```python |
| 139 | +qa = VectorDBQA.from_chain_type(llm=OpenAI(), chain_type="stuff", vectorstore=vectordb) |
| 140 | +``` |
| 141 | + |
| 142 | +To ask it a question, we can simply call the `invoke` method. |
| 143 | + |
| 144 | +```python |
| 145 | +query = "What are the top prize categories?" |
| 146 | +out = qa.invoke(query) |
| 147 | +print(out) |
| 148 | +``` |
| 149 | + |
| 150 | +This outputs: `The hackathon starts on Friday, February 16th at 9pm.` |
| 151 | + |
| 152 | +(You may notice a slightly different output when running this script, since ChatGPT is |
| 153 | +non-deterministic. It should be of similar content. however.) |
| 154 | + |
| 155 | +## Step 3 (optional): User Interface |
| 156 | + |
| 157 | +We have produced a minimally viable instance of RAG! However, most users are |
| 158 | +probably in want of a more friendly user interface for development or usage |
| 159 | +purposes. |
| 160 | + |
| 161 | +To achieve this, we can use the `gradio` library. |
| 162 | + |
| 163 | +```python |
| 164 | +import gradio as gr |
| 165 | +``` |
| 166 | + |
| 167 | +Gradio gives us a convenient chatbot template we can simply define some logic |
| 168 | +for. Let us first declare our chatbot response function. |
| 169 | + |
| 170 | +```python |
| 171 | +def response(message, history): |
| 172 | + h = '' |
| 173 | + |
| 174 | + for d in history: |
| 175 | + h += 'User message: \'' + d[0] + '\', ' |
| 176 | + h += 'Bot message: \'' + d[1] + '\' \n' |
| 177 | + |
| 178 | + m = 'You are an chatbot meant to answer participant questions about TreeHacks, a hackathon. Here is the prior message history: \n' + h + '\nHere is the message you have just been given: ' + message |
| 179 | + yield qa.run(m) |
| 180 | +``` |
| 181 | + |
| 182 | +This function accepts two variables: the most recent message from the user and a |
| 183 | +history of previous messages. We format the chat history into a single string |
| 184 | +such that our chatbot is always aware of the conversation's whole context. |
| 185 | +Although re-formatting this string every function call is certainly not the most |
| 186 | +elegant or efficient approach, it will suffice for our proof-of-concept. |
| 187 | + |
| 188 | +Notice that we also use this formatting step to provide additional context |
| 189 | +regarding the chatbot's purpose. This is a simple technique for focusing the |
| 190 | +chatbot's responses. |
| 191 | + |
| 192 | +To start our user interface, we can run the following line. |
| 193 | + |
| 194 | +```python |
| 195 | +gr.ChatInterface(response).launch() |
| 196 | +``` |
| 197 | + |
| 198 | +You should see a local URL printed in the terminal. Use this to access the gradio chat |
| 199 | +interface. |
| 200 | + |
| 201 | +# Thanks |
| 202 | +Thanks to everyone at Chroma for supporting this hackpack and TreeHacks! |
| 203 | + |
| 204 | +# Additional Resources |
| 205 | +- This hackpack is heavily derived from Harrison Chase's |
| 206 | + [chroma-langchain](https://github.com/hwchase17/chroma-langchain) demo repo. |
| 207 | + Please check it out! |
| 208 | +- Chroma has a variety of integratons and features, including multi-modal |
| 209 | + capabilities. Check out their [documentation](https://docs.trychroma.com/) to |
| 210 | + learn more. |
0 commit comments