Companies lose 42% of their knowledge when senior employees leave.
NEXUS turns your documents into an institution that remembers — answering with cited sources, grounding every claim in the evidence, and catching the moment two documents disagree.
The pipeline
From documents to defensible answers
Every query runs through four stages — and every stage is auditable.
1 · Ingest & chunk
PDF, DOCX, TXT parsed via Unstructured, split into 800-token semantic chunks with overlap and section headers.
Unstructured · semantic chunking2 · Hybrid retrieve
Dense pgvector search and sparse BM25 run in parallel, then fuse with Reciprocal Rank Fusion (k=60) — beating vector-only on keyword-heavy queries.
BM25 + pgvector · RRF3 · Generate
Groq GPT-OSS 120B streams the answer token-by-token over an LCEL chain — context wrapped in document tags to neutralise prompt injection. Natively bilingual: English and French.
Groq · GPT-OSS 120B · LCEL · SSE streaming4 · Verify
Auto-summary captures 4 key bullet points from each uploaded document instantly, while a second LLM pass compares retrieved chunks — surfacing contradictions and knowledge gaps before you trust the answer.
Auto-Summary · Contradiction RadarWhat NEXUS detects
Beyond chat-with-docs
Four capabilities that make answers trustworthy — not just fluent.
Contradiction Radar
A second structured LLM pass compares retrieved chunks across documents and flags conflicting statements with a severity rating.
Knowledge Gaps
Flags when the corpus has no real answer instead of hallucinating one — telling you what your documents are missing.
Source Attribution
Every answer cites its exact document, page, and excerpt — expandable inline, never a black box.
Auto-Summary
Instantly generates 4 key bullet points when you upload a document — capturing the most important facts, policies, and data in seconds.
See it think
It tells you when your own docs disagree
Four capabilities cycling live — summarize, cite sources, detect contradictions, and surface gaps.
Give your documents a memory.
One click loads a live corpus with built-in contradictions. No signup.