RAGs

Build.a.RAGs.bot.powered.by.LlamaIndex.-.26.November.2023.mp4

RAGs is a Streamlit app that lets you create a RAG pipeline from a data source using natural language.

You get to do the following:

1. Describe your task (e.g. "load this web page") and the parameters you want from your RAG systems (e.g. "i want to retrieve X number of docs")
2. Go into the config view and view/alter generated parameters (top-k, summarization, etc.) as needed.
3. Query the RAG agent over data with your questions.

This project is inspired by GPTs, launched by OpenAI.

Installation and Setup

Clone this project, go into the rags project folder. We recommend creating a virtual env for dependencies (python3 -m venv .venv).

poetry install --with dev

By default, we use OpenAI for both the builder agent as well as the generated RAG agent. Add .streamlit/secrets.toml in the home folder.

Then put the following:

openai_key = "<openai_key>"

Then run the app from the "home page" file.

streamlit run 1_🏠_Home.py

NOTE: If you've upgraded the version of RAGs, and you're running into issues on launch, you may need to delete the cache folder in your home directory (we may have introduced breaking changes in the stored data structure between versions).

Detailed Overview

The app contains the following sections, corresponding to the steps listed above.

1. 🏠 Home Page

This is the section where you build a RAG pipeline by instructing the "builder agent". Typically to setup a RAG pipeline you need the following components:

1. Describe the dataset. Currently we support either a single local file or a web page. We're open to suggestions here!
2. Describe the task. Concretely this description will be used to initialize the "system prompt" of the LLM powering the RAG pipeline.
3. Define the typical parameters for a RAG setup. See the below section for the list of parameters.

2. ⚙️ RAG Config

This section contains the RAG parameters, generated by the "builder agent" in the previous section. In this section, you have a UI showcasing the generated parameters and have full freedom to manually edit/change them as necessary.

Currently the set of parameters is as follows:

• System Prompt
• Include Summarization: whether to also add a summarization tool (instead of only doing top-k retrieval.)
• Top-K
• Chunk Size
• Embed Model
• LLM

If you manually change parameters, you can press the "Update Agent" button in order to update the agent.

If you don't see the `Update Agent` button, that's because you haven't created the agent yet. Please go to the previous "Home" page and complete the setup process.

We can always add more parameters to make this more "advanced" 🛠️, but thought this would be a good place to start.

3. Generated RAG Agent

Once your RAG agent is created, you have access to this page.

This is a standard chatbot interface where you can query the RAG agent and it will answer questions over your data.

It will be able to pick the right RAG tools (either top-k vector search or optionally summarization) in order to fulfill the query.

Supported LLMs and Embeddings

Builder Agent

By default the builder agent uses OpenAI. This is defined in the core/builder_config.py file.

You can customize this to whatever LLM you want (an example is provided for Anthropic).

Note that GPT-4 variants will give the most reliable results in terms of actually constructing an agent (we couldn't get Claude to work).

Generated RAG Agent

You can set the configuration either through natural language or manually for both the embedding model and LLM.

• LLM: We support the following LLMs, but you need to explicitly specify the ID to the builder agent.
  • OpenAI: ID is "openai:<model_name>" e.g. "openai:gpt-4-1106-preview"
  • Anthropic: ID is "anthropic:<model_name>" e.g. "anthropic:claude-2"
  • Replicate: ID is "replicate:<model_name>"
  • HuggingFace: ID is "local:<model_name>" e.g. "local:BAAI/bge-small-en"
• Embeddings: Supports text-embedding-ada-002 by default, but also supports Hugging Face models. To use a hugging face model simply prepend with local, e.g. local:BAAI/bge-small-en.

Resources

Running into issues? Please file a GitHub issue or join our Discord.

This app was built with LlamaIndex Python.

See our launch blog post here.

Example prompts for this specific agent

Here are some sample prompts you can use to interact with the agent "Alphabet_Earnings_Calls_23 Q1-Q3":

1. "Provide a financial summary of the Q1 2023 earnings call."

System Prompt: "Utilize the available analytical tools to generate a comprehensive financial summary of the Q1 2023 earnings call. Ensure that all insights, figures, and conclusions are derived using these tools. Present the summary in a clear and concise manner."

2. "What are the key growth drivers mentioned in the Q2 2023 earnings call?"

System Prompt: "Please utilize the available analytical tools to identify and summarize the key growth drivers discussed in the Q2 2023 earnings call. Ensure that you process the information thoroughly using the tools before providing an answer."

3. "Can you forecast the potential revenue trends for Q4 2023 based on the data from Q1 to Q3?"

System Prompt: "Using the analytical tools at your disposal, please generate a forecast of the potential revenue trends for Q4 2023. Ensure that your analysis is based on the data available from Q1 to Q3 of the same year. Your response should reflect a comprehensive analysis and include any relevant patterns or insights that can be derived from the data for the first three quarters."

Follow up: What else can you tell me about it?

System Prompt: "provide additional insights on Q4 2023 revenue trends based on Q1 to Q3 data

4. "What is the key factor for the growth in Youtube talked about in the Q1 earnings call?"

System Prompt: ?

What can we learn from these example prompts?

Keeping in mind that the format we used for the example prompts mostly gives us good and usable results, we can say that the combination of the prompt layout and the parameter setup work well together. When looking at the prompts more in depth, we see that they always specify one earning call for the result (Metadata Filters: https://docs.google.com/presentation/d/1IJ1bpoLmHfFzKM3Ef6OoWGwvrwDwLV7EcoOHxLZzizE/edit#slide=id.g1e95adadc30_0_1239), select a distinct topic while using specific keywords. It should be noted that the user needs general knowledge about the source's content so a sensible keyword for the topic can be selected.

When we create new custom RAG agents, we should supply the user with key metadata tags to make sure their prompts work.