of company knowledge walks out the door when senior employees leave. nexus doesn't let that happen.
Upload your documents in any of 15 formats. Get an AI summary instantly. Ask anything — every answer is traceable to a source, contradictions are surfaced automatically, and gaps are declared instead of hallucinated.
BM25 keyword search fused with pgvector semantic search via Reciprocal Rank Fusion. MATCH_K=8, threshold=0.40. Catches what pure vector search misses — and what pure keyword search can't semantify.
⚡02
Auto-Summary on Upload
The moment a document lands, one LLM call returns a one-liner, 4 insight bullets, and 3 clickable question chips that fire directly into chat. Zero friction from upload to first answer.
🔴03
Contradiction Radar
A structured second LLM call compares retrieved chunks across documents. When policies diverge, it surfaces both statements side-by-side with severity rating and explanation — never silently picks one.
🕳️04
Knowledge Gap Detection
If coverage score falls below 0.45 and the query carries a non-answer signal, nexus declares a knowledge gap instead of hallucinating. The system knows what it doesn't know.
ARCHITECTURE
The retrieval pipeline, step by step.
Upload
15 formats
→
Chunk
800 tokens
→
Embed
all-MiniLM
→
BM25+vec
pgvector
→
RRF
rerank
→
LLM
Groq
→
Ground
claim-level
Every chunk is embedded with all-MiniLM-L6-v2 (CPU, baked into Docker) · BM25 index rebuilt on each upload · RRF(k=60) fuses both scores
CONTRADICTION RADAR
When documents disagree, nexus says so.
A structured second LLM call compares retrieved chunks. Both versions surface — no silent overwrite.
Q:"What's the remote-work policy? How many days per week?"
TechCorp_HR_Policy_2023.pdf · p.1
"Employees may work remotely for up to 3 days per week, subject to manager approval."
TechCorp_HR_Policy_2024.pdf · p.1
"The updated remote-work allowance is 2 days per week, effective January 2024."
CONTRADICTION DETECTED · medium severity
The 2023 and 2024 HR policies diverge on remote-work limits. The 2024 version appears more recent and may supersede the 2023 policy — but both are present in your corpus and nexus will not silently pick one.
CONFIDENCE GATING
High confidence answers. Low confidence refusals.
Every answer is grounded via dual-path verification: citation-aware fuzzy matching + semantic cosine against embedded chunks. Coverage below 0.45 triggers a knowledge gap declaration. No hallucination by design.