This repository documents the architecture and evaluation plan for ReasonedAI, a private academic GraphRAG project. It contains documentation only: no application source code, private report, credentials, private URLs, user data, deployed endpoint, or current benchmark results.
flowchart LR
U[Authenticated user] --> W[React web client]
W --> API[FastAPI service]
API --> G[Graph + vector retrieval]
G --> N[Neo4j Aura]
API --> V[Vertex model services]
V --> X[Extraction / embeddings / answer / verifier]
X --> C[Citations + grounded response]
- React frontend and FastAPI backend.
- The documented design restricts authentication and document access to approved users.
- Google Cloud Storage for document/object handling.
- Neo4j Aura for graph, vector, and provenance relationships.
- Vertex model services for extraction, embeddings, answer generation, and verification.
- Hybrid retrieval: vector similarity followed by bounded graph traversal and provenance-aware citations.
These files describe the system design. They do not provide a runnable application, SDK, hosted service, or benchmark implementation.
Vector retrieval is useful for semantic recall; a graph adds explicit entity relationships, constrained traversal, and provenance links. The intended answer path is therefore: retrieve candidate evidence, traverse only bounded relevant relationships, generate from cited context, and run a separate groundedness check before returning the response.
The architecture is derived from the project's private final report. Published files are limited to architecture, evaluation, limitations, and security notes. The application source code, report, screenshots, identity configuration, dataset, credentials, cloud exports, live endpoint, and current benchmark results are not available in this repository. See LIMITATIONS.md for details.