RAG Document Assistant

Overview

A Retrieval-Augmented Generation (RAGRAGRetrieval-Augmented Generation — fetching relevant passages from a document store at query time and handing them to a language model as context, so its answer is grounded in those sources.) application built for the Endee.io Campus Recruitment Assignment. Users upload text documents, which are stored as vector embeddingsembeddingA list of numbers (a vector) representing a piece of text, an image or audio so that similar meanings land close together — the basis of semantic search. in the Endee Vector Databasevector databaseA database built to store embeddings and run fast similarity search over them, usually with an approximate nearest-neighbour index.; natural-language questions then retrieve the most semantically relevant chunks and feed them as grounded context to a Groq-hosted LLMLLMLarge language model — a neural network trained on large amounts of text to predict the next token, used for generation, summarisation, classification and more. (llama-3.3-70b-versatile). Answers are current, domain-specific, and traceable to source documents — the LLM never operates without retrieval grounding.

Architecture

• Ingestion pipeline — documents split into configurable 500-word chunk windows, each independently embedded with all-MiniLM-L6-v2 (384-dim float32) and batch-inserted into Endee
• HNSWHNSWHierarchical Navigable Small World — a graph-based index for approximate nearest-neighbour search over vectors. A query hops through layered proximity graphs instead of comparing against every vector. vector index — cosine-similarity approximate nearest neighbor search (M=16, ef_con=200) over MDBX persistent storage, returning top-k results in milliseconds
• Query pipeline — the question is encoded into the same vector space, top-k chunks retrieved via Endee's search API (MessagePack-encoded responses), assembled into a context prompt for Groq
• Source attribution — every answer renders the exact retrieved chunks with similarity scores
• Auto-recovery — detects Endee's "required files missing for index" error on ephemeral hosts, transparently deletes the stale index, reprovisions, and retries ingestion
• Defensive parsing — a multi-format metadata parser handles raw dicts, JSON strings, and bytes across Endee versions
• Streamlit frontend with dark theming, real-time system status, live metrics (docs / chunks / latency), and configurable top-k, temperature, and chunk size
• Render blueprint deploy — render.yaml provisions both the Dockerized Endee server and the Streamlit app with internal-hostname networking

What It Demonstrates

The full RAG lifecycle end to end: chunking strategy trade-offs, symmetric query/document embedding, ANN index configuration, binary protocol handling (msgpack), prompt construction for grounded generation, and resilience engineering for free-tier cloud infrastructure.