Citizen Advice RAG Flow Documentation

Overview

This Langflow flow implements a Retrieval Augmented Generation (RAG) system designed to answer questions based on content from the Citizen Advice website. It loads and ingests website data, processes and indexes it into a vector database, and then uses a language model to answer user queries with relevant, factual information.

Flow Components and How They Work Together

1. Data Loading and Ingestion

URL Component

• Recursively crawls and loads web pages starting from the Citizen Advice benefits URL (https://www.citizensadvice.org.uk/benefits/)
• Configurable max depth (default 2) controls how many link levels are followed
• Option to restrict crawling to the same domain to avoid external sites
• Supports asynchronous loading for faster crawling
• Outputs raw text data extracted from HTML pages as a list of documents

Split Text Component

• Splits the loaded text documents into smaller chunks for better embedding and retrieval
• Configured with chunk size (default 1000 characters) and overlap (default 200 characters) to maintain context between chunks
• Outputs chunks as data objects

MistralAI Embeddings Component

• Converts text chunks into vector embeddings using MistralAI embedding models
• Requires a Mistral API key for authentication

Chroma Vector Store Component

• Stores embeddings and associated text chunks in a Chroma vector database for efficient similarity search
• Configured with collection name (cas-benefits) and persistence directory (./chroma-db)
• Supports options like allowing duplicates and caching

2. Query Handling and Prompt Construction

Chat Input Component

• Entry point for user queries entered in the Langflow Playground interface
• Captures user messages and metadata such as sender info and session ID

Chroma Vector Store (Search Query Input)

• Receives the user query and performs a similarity search in the vector store to retrieve relevant documents
• Outputs search results as data objects

Parser Component

• Formats the retrieved search results into a textual context using a template (e.g., {source} or {text})
• Supports cleaning data and converting structured data into readable text

Prompt Component

• Combines the retrieved context, source URLs, and user question into a prompt template
• The prompt instructs the language model to answer factually based on the retrieved content and to include source URLs without duplicates

3. Language Model and Output

MistralAI Model Component

• Generates the answer text using MistralAI large language models
• Configurable parameters include model name, temperature, max tokens, API key, and more
• Input is the prompt message constructed by the Prompt Component
• Output is the generated answer as a message

Chat Output Component

• Displays the generated answer in the Langflow Playground chat interface
• Supports storing the message in chat history and customizing sender info and appearance

4. Additional Components and Notes

Batch Run Component

• Allows running the language model over each row of a DataFrame, useful for batch processing or evaluation

Correctness Evaluator Component

• Uses OpenAI's LLM to evaluate the correctness of model outputs against reference answers

Privacy Anonymizer Component (Optional)

• Can anonymize input text by masking sensitive information (PII), useful for privacy compliance

Note Nodes

• Provide inline documentation and guidance within the flow workspace, such as reminders to add API keys and instructions on running the data loading flow first

How to Build a Similar Flow in Langflow

1. Load Data

• Use a URL loader to crawl and extract text from your target website or data source
• Split the text into chunks suitable for embedding

2. Generate Embeddings

• Use an embedding model component (e.g., MistralAI Embeddings) to convert text chunks into vectors

3. Create a Vector Store

• Use a vector store component (e.g., Chroma) to index and store embeddings for similarity search

4. Set Up Query Input and Retrieval

• Use a Chat Input component to accept user queries
• Connect it to the vector store to perform similarity search and retrieve relevant documents

5. Format Retrieved Data

• Use a Parser component to convert retrieved documents into a readable context string

6. Build Prompt

• Use a Prompt component to combine context, sources, and user question into a prompt template

7. Generate Answer

• Use a language model component (e.g., MistralAI) to generate answers from the prompt

8. Display Output

• Use a Chat Output component to show the generated answer in the chat interface

9. Add Optional Components

• Add batch processing or evaluation components to test and improve your system
• Add privacy components if needed

10. Document Your Flow

• Use Note nodes to add helpful comments and instructions for users