Looking at LangFlow
Riding the wave of Gen AI are tools that aim to make it simple to build RAG (Retrieval Augmented Generation) flows. LangChain is a popular framework that has a concept of “chains” to orchestrate the flows. The flows typically have the following steps:
1. Load a document
2. Split it into chunks
3. Store the chunks into a vector store such as Chroma
4. Retrieve the relevant chunks
5. Send the chunks along with prompt to the LLM and get the response
RAG helps in supplementing contextual information and getting responses from the LLM with better accuracy. The chunking step helps in retrieving content with similar semantics. LangChain has many components to help build the chain including but not limited to
- Document loaders
- Text splitters for chunking
- Several helpers to work with vector stores and embeddings
- Prompt builders
Typically, coding is done using Python but JavaScript is also supported.
LangFlow is a browser based tool built on top of LangChain and provides a visual approach to development. There are pre-built templates available that can be used as starting points. These are impressive as they capture almost the entire functionality needed to build a vanilla flow. So much so that the flows can be run just by providing the inputs and token keys.
The LangFlow interface is intuitive. It has a palette on the left hand side with the components to get input, output, handle data, helpers etc. The canvas is already filled with the relevant components for the RAG flow. In the flow, there are two parts, the top part handles the chat input which could have the question, the vector store component using Astra DB which is from DataStax itself, building of prompts and producing the output. The bottom part contains the chunking process into Astra DB from the file loaded.
The easiest way to try out LangFlow is by signing up for the DataStax version on the cloud or through Hugging Face Spaces. If you prefer to work locally, a Python pip version too is available.
Though it is a 1.0 release, there are a few quirks that I noticed, such as directory access for files is not seamless. It requires custom code to read. As it evolves, the simplicity and ease of use could make LangFlow a tool of choice to build RAG flows.
Cross posted on My LinkedIn account
