CRAG

Corrective-RAG (CRAG) is a strategy for RAG that incorporates self-reflection / self-grading on retrieved documents.

Here in this project, I have built a CRAG model using LangGraph, Tavily Search API and ChatFireworks API.

Workflow:

• Upload your pdf files.
• They will be converted into embeddings using GPT4Embeddings(model_name="all-MiniLM-L6-v2.gguf2.f16.gguf") and will be stored in Chroma vectordatabase.
• The user will ask a quetion.
• The relevant documents are retrieved.
• The documents are graded as "yes" that are relevant to the question or "no" not relevant to the question using the LLM of ChatFireworks.
• If "yes" the documents are then used for final generation of the output.
• If "no",the question will be transformed and the tavily search api is used to extract relevant data from the internet.
• The final output will be generated using the extracted documents from the internet.

The final model contains 6 main nodes as shown in the figure.

API's used

• Fireworks API
• Tavily Search API