NexusScholar

NexusScholar is VTG’s document-intelligence layer — a retrieval-augmented generation (RAG) capability that turns large PDF and document collections into a searchable knowledge base and answers natural-language questions with grounded, cited evidence. It combines semantic embedding, vector search and cross-encoder reranking with local large-language-model generation, and adds a prompt-driven toolkit to summarise, expand and reshape source text.

Built on the NexusFlow reasoning layer and assurable via NexusTrust, NexusScholar runs fully on-premise for data-sovereign deployments — no data leaves your infrastructure.

---

Key Features

• Semantic indexing of large PDF collections with a per-document embedding cache
• Retrieval with cross-encoder reranking for high-precision, grounded answers
• Page-level source citations on every answer for full traceability
• Prompt-driven toolkit — summarise, expand, simplify, abstract and rewrite
• Map-reduce summarisation of whole documents beyond the model’s context window
• Fully local, on-premise operation — no data leaves the deployment

---

Application Areas

---

Technical Specification

Architecture	Modular RAG pipeline: ingest → embed → index → retrieve → rerank → generate
Models	Local embedding + cross-encoder reranker with an Ollama-served LLM (swappable)
Compute	CPU-friendly; optional GPU accelerates embedding and generation
Integration	REST API with token-streaming (SSE) endpoints; built on NexusFlow
Deployment	Fully on-premise and offline-capable; assurable via NexusTrust for oversight
Clusters	Cross-cutting — Clusters 1, 2, 3, 4, 5, 6

---

How It Works

NexusScholar processes document collections through a six-stage pipeline designed for accuracy and traceability:

1. Ingest

PDFs and documents are parsed, cleaned, and split into semantically meaningful chunks with metadata preserved — page numbers, section headings, and document identifiers.

2. Embed & Index

Each chunk is transformed into a dense vector embedding and stored in a per-document cache. The vector index enables fast approximate nearest-neighbour search across the entire collection.

3. Retrieve

When a question is asked, the system retrieves the most semantically relevant passages from the vector index — casting a wide net to ensure no critical evidence is missed.

4. Rerank

A cross-encoder reranker re-scores the retrieved passages against the original question, pushing the most precisely relevant evidence to the top — significantly improving answer quality over embedding-only retrieval.

5. Generate

The top-ranked passages are provided as context to a locally served LLM, which generates a natural-language answer grounded entirely in the source material — with page-level citations for every claim.

6. Transform

The prompt-driven toolkit enables further operations on the answer or source text: summarise, expand, simplify, abstract, or rewrite — giving users flexible control over the output format and depth.

---

Heritage & Lineage

Developed by VTG as the document-intelligence layer of the Nexus suite, generalising retrieval-augmented generation over research and technical document collections. NexusScholar is built on the NexusFlow reasoning core and assured through NexusTrust.

Target Calls: Horizon-wide horizontal capability — applicable to Cluster 1, 2, 3, 4, 5 and 6 calls requiring evidence synthesis, document intelligence, or trustworthy AI-assisted knowledge work.

---

Try NexusScholar

NexusScholar is available as a hosted service for evaluation and demonstration purposes. Upload your document collections and start asking questions with grounded, cited answers.