NOUSMultimodal intelligence platform
Grounded insights, with citations you can trust
An AI research assistant that reads, reasons, and writes. NOUS ingests any modality, retrieves across a hybrid index, and links every claim to a source chunk, so nothing is generated without a reference.
- Inline citations on every claim
- Human-in-the-loop on every write
- Sources cited
- 4
- Ungrounded claims
- 0
No single tool retrieves, reasons, and cites.
Knowledge workers drown in unstructured information across disconnected tools, and the usual options each fall short in a different way.
- Search
- Returns keywords, not understanding. You still read everything.
- LLM chat
- Reasons fluently, then hallucinates without a source you can check.
- Note apps
- Store what you write. They never connect or reason across it.
NOUS is one grounded workspace that does all three, and shows its work.
What NOUS does
A research-grade pipeline, from ingest to a cited draft you can export.
Agent system, with a human in the loop
An intent-routed LangGraph StateGraph with specialized subgraphs for research, writing, and data. Every destructive tool is gated behind interrupt(), so the agent surfaces the pending action and waits. Checkpointed in Postgres, so threads survive restarts.
Hybrid retrieval
Keyword search (Postgres FTS) and graph traversal (Neo4j), fused and reranked with Cohere. One query hits text chunks, image regions, audio segments, and citation edges together.
Citation pipeline
Hybrid extraction from arXiv, Semantic Scholar, CrossRef, and PDF parsers into an interactive Neo4j citation graph. Export to BibTeX, APA, IEEE, or MLA, with warnings on incomplete metadata.
Multi-agent literature reviews
A writing subgraph runs theme extraction, per-theme synthesis, and inline citation injection, then exports LaTeX and a .bib file. Versioned, so you can iterate on a draft.
Multimodal ingest
PDF with OCR, image object and scene detection, Whisper speech-to-text with speaker diarization, and video frame plus audio extraction. One pipeline, one index.
Local AI in the browser
WebLLM runs Llama, Phi, Gemma, and Qwen for offline or privacy-sensitive workflows. No data leaves the client.
Trust is measured, not claimed
Quality is enforced with DeepEval thresholds in CI. A regression past any of these blocks the merge, so grounding does not drift over time.
- >90%
- Faithfulness
- claims backed by sources
- >70%
- Answer relevancy
- on grounded evals
- <10%
- Hallucination
- regressions block merge
One graph, end to end
A checkpointed StateGraph routes intent to the right subgraph, runs tools, and saves memory. The interrupt and resume cycle keeps a human in control of every write.
- rag
- intent
- memory
- route
- tools
- save
Built on
- Next.js 15
- FastAPI
- LangGraph
- Neo4j
- Postgres
- Redis
- Cohere rerank
- Whisper
Read, reason, and write with sources you can trust
Explore the live showcase, or read the source. NOUS is open and built in the open.