Open Source

Dual-mode RAG application demonstrating both traditional and multimodal RAG capabilities. Public mode leverages AWS Bedrock for cloud-based inference, while Private mode runs fully local with Ollama and Docling, ensuring no data ever leaves the machine.

• Two distinct pipelines: text extraction with layout-aware document processing (PDF, DOCX, TXT) via Docling, and direct multimodal image embedding via Nomic Vision and Cohere Embed v4
• Hybrid retrieval combining vector search, BM25, and metadata filtering with cross-modal retrieval enabling text queries against image collections in the same latent space

Autonomous ReAct agent that reasons about each query and picks the best retrieval tool at runtime. Unlike a fixed pipeline, the LLM drives its own decision loop (Reason, Act, Observe, Repeat) until it has enough context for a final answer.

• 4 agent tools (vector search, file search, file read, web search) with priority-based selection via engineered system prompt. No hard-coded routing logic.
• Live reasoning traces in Streamlit UI, MCP server with dual transport (stdio + SSE), local embeddings with cloud inference

Developer-controlled RAG pipeline using LangGraph state machines. Unlike the ReAct approach, the graph controls routing between 6 nodes with deterministic transitions and quality gates at every stage.

• Relevance grading + query rewriting with bounded retries (max 2), a self-correcting retrieval loop that degrades gracefully when retries are exhausted
• Hallucination detection as a post-generation gate, full execution trace visible in Streamlit UI for complete pipeline observability