Skip to content

Latest commit

 

History

History
92 lines (61 loc) · 4.33 KB

chainlit.md

File metadata and controls

92 lines (61 loc) · 4.33 KB
Raw

Welcome to Sage Chat! 🚀🤖

Sage Chat is an AI Assistant designed to interact with various data sources and provide answers to questions.

The Chat UI interface is powered by Chainlit. Learn more in the Chainlit Documentation. 📚

Chat Architecture

Sage Chat leverages the RAG system to provide accurate and reliable answers to user queries.

Retrieval-augmented generation (RAG) is a technique that enhances the capabilities of generative AI models by incorporating facts retrieved from external sources.

The core components of Sage Chat include:

  • LLMs (Large Language Models)
  • Embeddings
  • Sources
  • Prompts

Indexing Process

The sequence from raw data to answer involves the following steps:

  • Load: Data is loaded using loaders in the Sources Module.
  • Split: Text splitters break large documents into smaller chunks, facilitating indexing and model input.
  • Store: Chunks are stored and indexed using a VectorStore and Embeddings model, enabled by the FAISS Vector Database and a combination of Jina/OpenAI Embeddings.

Retrieval and Generation Process

  • Retrieve: Relevant document chunks are retrieved from storage in response to user input using a Retriever.
  • Generate: A ChatModel / LLM formulates an answer using a prompt that incorporates both the question and the retrieved data.

Chain Overview

The architecture of Sage Chat is depicted below:

Alt text

Key components include:

  • Condense Question LLM Chain
  • Retriever and Reranking Chain
  • Question & Sources plus Agent Executor Chain

Condense Question LLM Chain

This stage processes the user input question along with chat history to generate a new standalone question:

HUMAN:
question: "Team members in the Xerex team?"
history: []

AI:
{
    How many team members are in the Xerex team?"
}

Retriever and Reranking Chain

This stage involves converting the standalone question to a vector representation and performing a similarity search to retrieve related documents. If reranking is enabled, it further refines the search results to improve relevance.

Currently supported rerankers include:

  • Cohere Reranking (Proprietary)
  • BGE Reranker (Open-source)

Question & Sources plus Agent Executor Chain

This final stage varies based on the user-selected mode in the chat interface:

  • Chat-only mode (default): The results from reranking/retrieval are passed to an LLM chain to answer the question using the retrieved documents as a knowledge source.
  • Agent mode: The agent executor, equipped with tools, actively combines data from retrieval/reranking and any available tools to answer the question, possibly involving multiple steps.
  • File mode: In this mode, Sage is not configured with any retriever and it insteads creates a real-time retriever for user's upload documents and the user can then ask questions about the documents provided.

Tips

If Sage is unable to answer a question, even with the correct information in the sources, users can improve the response by providing more detailed questions or guidance, which can significantly enhance the results.

For instance - To force the condense question chain from not rephraseing the original question you can add some quidance

# -------Before--------------
HUMAN:
Question: "In your own words or possibility, how many years do humanaility have left to able to achieve AGI? I understand you don't know but I insist to give some estimates here - take a look at historical human advanment in the last few decades."
AI: "Based on the progress in AI over the past decades, what is an estimated timeline for the achievement of Artificial General Intelligence (AGI)?"
# -------After--------------
HUMAN:
Question: "In your own words or possibility, how many years do humanity have left to able to achieve AGI? I understand you don't know but I insist to give some estimates here - take a look at historical human advancement in the last few decades."

Return question as it is - DO NOT MODIFY PLEASE"
AI: "In your own words or possibility, how many years do humanity have left to able to achieve AGI? I understand you don't know but I insist to give some estimates here - take a look at historical human advancement in the last few decades."