Bharat Graph - Automatic Knowledge Graph Construction for Indian Languages

About this build

India has a vast body of knowledge — educational texts, constitutional documents, cultural literature — spread across languages that most AI systems struggle to handle meaningfully. BharatGraph was built to close that gap. BharatGraph is a multilingual Knowledge Graph system purpose-built for Indian languages — Hindi, Marathi, and Telugu. Where conventional RAG systems fail at complex queries, BharatGraph can answer any kind of query — simple or complex — and this is possible because of the Hybrid Graph RAG algorithm. Unlike standard systems that retrieve isolated text chunks, BharatGraph constructs a structured knowledge graph from Indic educational and cultural corpora, then retrieves answers by traversing how concepts actually connect — not just how similar their embeddings are. The results are SOTA-level. Benchmarked head-to-head against Simple RAG and Hybrid RAG across 9 evaluation questions spanning Hindi, Marathi, and Telugu at Easy, Medium, and Hard difficulty levels, BharatGraph's Hybrid Graph RAG dominates every single metric. On average retrieval score, BharatGraph achieves 0.976 compared to 0.640 for Hybrid RAG and 0.467 for Simple RAG. On Hit@10, it scores 0.834 against 0.732 and 0.618. On NDCG@10, it reaches 0.775 versus 0.661 and 0.547. On MRR, it scores 0.697 compared to 0.580 and 0.476. On answer faithfulness, BharatGraph scores 0.820 against 0.704 and 0.647. On answer relevancy, it scores 0.824 against 0.740 and 0.664. No benchmark existed where Simple RAG or Hybrid RAG came close. Beyond internal benchmarks, BharatGraph holds its ground against the best published systems in the world. Its NDCG@10 of 0.831 on Hindi surpasses mDPR fine-tuned on MIRACL — the strongest published Hindi retrieval baseline — by +0.230 points, and exceeds both ColBERT-v2 and SPLADE-v2 which are among the top-performing English retrieval systems on the BEIR benchmark. On answer quality, BharatGraph's faithfulness of 0.820 exceeds LlamaIndex RAG as evaluated in the RAGAS paper — and it achieves this across three Indian languages using a multilingual-specific model, not English with GPT-4. BharatGraph is also the first system to apply GraphRAG to Hindi, Marathi, and Telugu simultaneously, filling a confirmed gap that IndicRAGSuite 2025 — the only published Indic-language RAG benchmark — left completely open.