RAG-LCC
An experimental retrieval-augmented generation lab for studying why RAG pipelines fail under constrained context windows, modest hardware and multi-turn chat, treating document classification, chunking and retrieval strategy as tunable, comparable parameters.
RAG-LCC, short for Retrieval-Augmented Generation Local Corpus and Classification, is an experimental lab for studying and controlling retrieval and context assembly under real-world constraints: limited context windows, modest GPUs, large documents and multi-turn chat. Rather than pushing toward ever-larger context windows, it treats document classification, chunking, retrieval strategy and content filtering as first-class, tunable architectural decisions, aimed at the failure modes where standard embed-and-cosine-similarity RAG hallucinates on conflicting information, drifts across a multi-turn conversation, or fails silently on minority-language documents.
Its components include a document classifier whose output can filter what its loader ingests, a loader for text extraction from document formats, images and MS Office files with vector database ingestion, and a chat component that integrates with Open WebUI. It is meant as an insight tool for experimenting with the parameters that affect retrieval quality on a small test corpus, rather than an everyday production RAG system, and runs against Ollama and vLLM in Dev Containers or natively on Linux and Windows.
It is written in Python and MIT licensed.