ReliSource Protocol Parser: A Private, Local-First RAG Platform for Clinical-Trial Protocols
Clinical-trial protocols are long, dense, and highly structured — and reviewers spend hours hunting through them for specific requirements. The ReliSource Protocol Parser (RPP) is a local-first GenAI/RAG platform that turns those documents into a searchable knowledge base you can query in plain language and get back cited, verifiable answers.
The core constraint: privacy
Clinical protocol data is sensitive and often can't leave a controlled environment. That single requirement shaped the whole architecture: the retrieval and extraction stack is designed to run inside a private environment, with a local LLM option so that sensitive text never has to be sent to a third-party service.
How the pipeline works
Ingest & chunk: protocol documents are parsed and split into passages suitable for retrieval.
Embed: each passage is converted into a vector embedding and stored in a vector database.
Retrieve (hybrid): a query is matched using hybrid search — dense semantic vector search combined with keyword search — to surface the most relevant passages.
Extract & answer: an LLM reads the retrieved passages and produces an answer with citations back to the source text.
Validate & export: a human-in-the-loop step lets reviewers confirm results, which can then be exported in structured form.
Technologies used
Large language models (hosted & local): for extraction and question-answering; a local model keeps sensitive data in-boundary.
Embeddings + vector database: encode passages as vectors and retrieve them by semantic similarity.
Hybrid search: combines dense (vector) and sparse (keyword) retrieval for better recall on domain terms.
LangChain: orchestrates the retrieval-augmented-generation flow.
LangSmith: traces and evaluates the pipeline so answer quality can be measured and debugged.
Human-in-the-loop validation, citations & exports: keep outputs trustworthy and auditable, with mobile-ready search workflows.
Impact
RPP cuts the manual effort of protocol review while keeping every piece of sensitive clinical data inside a private, controlled environment — and every answer is traceable back to its source.
Tech stack: Python, LangChain, LangSmith, hosted & local LLMs, text embeddings, vector database, hybrid (dense + sparse) search, retrieval-augmented generation (RAG).
Comments