Sarvam AI · Nov 2023 – Jan 2024
An end-to-end CSV QA agent benchmarked across GPT-3.5/4, Llama and Mistral, a multilingual OCR pipeline for non-UTF-8 Indic PDFs, and RAG over FAISS and Pinecone with RAGAS-based evaluation. The visualisation below follows one query from document scan to scored answer.
01 The scaffold
A document holding Devanagari, Tamil and Telugu text sits beside a vector space of embedded chunks with one query vector, an LLM running a ReAct loop of Thought, Action and Observation, and a RAGAS gauge at zero. Everything is laid out; nothing has run yet.
02 The run
The OCR scan sweeps the page and latches a bounding box onto each script. The query's top-k neighbours light up with their cosine ranking, the retrieved context drives the ReAct loop around to an answer for the CSV question, and the RAGAS gauge ticks up through faithfulness and answer relevance.
The pipeline drawn above: OCR with bounding boxes over non-UTF-8 Indic scripts, chunks embedded into a vector store, and approximate nearest neighbour search in FAISS or Pinecone returning the top-k context by cosine similarity. The context is injected into an LLM running a ReAct loop, alternating Thought, Action and Observation until it answers the CSV aggregation question, and the output is scored with RAGAS on faithfulness and answer relevance.
The first scroll lays out the scaffold. The second animates the run: the scan line, the retrieval, the reasoning loop and the evaluation, each a step you can point to when the agent's answer needs to be trusted.