Knowledge Graph RAG Assistant

HackerEarth asked the team to explore whether dense retrieval and a knowledge graph could supply complementary context to a retrieval-augmented generation system. We built a React and FastAPI application that routed a question through language processing, DBpedia lookups, FAISS search, and answer synthesis.

The retrieval corpus combined 10,000 Wikipedia articles with custom class-note PDFs. The final handoff included Docker and Docker Compose configuration, automated tests, GitHub Actions, and system documentation.

My contribution centered on retrieval: I implemented text embeddings with SentenceTransformers, built the FAISS vector-search path, and contributed to the documentation used for delivery and maintenance.

The maintained public fork now packages a deterministic demo path and an evidence-forward interface, making the architecture reviewable even when the original API credentials and multi-gigabyte corpus are absent.

The chat handler coordinated two retrieval branches. The knowledge-graph handler queried structured entity relationships, while the vector-search handler embedded the question and retrieved semantically similar passages. Their context converged at the language-model handler for final response generation.

An early end-to-end query took roughly 45 seconds. A model change and pipeline optimizations brought observed responses into the 6–8 second range, with vector search remaining the dominant cost in the final measurement.

• Response relevance was assessed primarily through manual and client testing rather than a formal retrieval benchmark.
• Entity extraction was less reliable for complex questions involving multiple entities or relationships.
• The complete system depended on DBpedia, an LLM API, and local corpus artifacts that are not bundled with the public repository.
• A 7.6-second average response remained noticeable for an interactive assistant and left room for retrieval and caching work.