ToS RAG Analyzer
A Retrieval-Augmented Generation system that makes Terms of Service and privacy policies transparent. Every answer is grounded in real policy text through a vector retrieval pipeline, so plain-English explanations always trace back to the clauses they came from.
2025
AI Systems Engineer
- Python
- Streamlit
- FAISS
- Hugging Face
Tech Stack
Technologies and tools used to bring this project to life.
- Python
- Streamlit
- FAISS
- Hugging Face
- Sentence Transformers (all-MiniLM-L6-v2)
- OpenAI GPT-3.5
- OPP-115
- RAG
My Role
AI Systems Engineer • A few of the surfaces I shaped on this project
Vector pipeline and retrieval
Policy chunks are embedded with Hugging Face all-MiniLM-L6-v2 and indexed with FAISS, so each user query is matched against semantic vectors rather than keyword overlap. The result is sub-second top-k retrieval that stays accurate across phrasing variations.
Common-concerns templates
A curated set of the questions users actually have about a policy (data sharing, retention, third parties, opt-outs) gives them a one-click way to interrogate any document without having to formulate the query from scratch.
Grounded answers with citations
Every response surfaces the plain-English answer alongside the policy chunks it was derived from. Because generation is constrained to the retrieved context, the citations are not decorative; they are the only source the LLM is allowed to reason from.
Pluggable generation backends
The same retrieved context can be routed through GPT-3.5 or through an open-source Hugging Face model, making the system useful both for higher-quality answers and for fully self-hosted, cost-free deployments.
Validated on live policies
Tested end-to-end against real-world Terms of Service the system was not pre-loaded with (UberEats among others) to confirm that retrieval quality and grounding hold up when the corpus is swapped at runtime.
The ToS RAG Analyzer is a retrieval-augmented AI system designed to make Terms of Service and privacy policies easier to understand through grounded, citation-backed answers.
I developed the full RAG pipeline, including document chunking, semantic retrieval, vector indexing, and constrained generation workflows. A key challenge was reducing hallucination while keeping responses understandable for non-technical users.
The system uses FAISS vector search, Hugging Face sentence transformers, and interchangeable GPT/Open-source generation backends, all deployed through a Streamlit interface.