Upload your knowledge base, define your objective, and let NEXUS run 7 specialised AI agents in parallel — each grounded in your documents via Pinecone Hybrid RAG — delivering a complete strategic brief in under 30 seconds.
All 7 chains fire simultaneously via LangChain LCEL's RunnableParallel. Each agent retrieves its own document context from Pinecone — your full strategy brief is ready in seconds, not hours.
A five-stage async pipeline — knowledge base indexing handled separately by Celery so the generation pipeline never blocks on I/O.
Upload PDFs, DOCX, or TXT files. A Celery background task chunks each document and indexes it into Pinecone Serverless using BAAI/bge-small-en-v1.5 embeddings (384-dim). Indexed once, reused forever.
Celery · Pinecone · HuggingFaceChoose which KB documents to ground the analysis in. Define your strategic objective. Document IDs are passed directly — no re-indexing, no file uploads at pipeline time.
Django REST · JWT AuthEach of the 7 agents runs its own semantic query against Pinecone. Results are re-ranked by keyword overlap + cosine similarity. Every agent gets context tailored to its analysis domain.
Pinecone · Hybrid RAG · Re-rankRunnableParallel fires all 7 chains at once via Celery. Pipeline events — start, RAG done, each agent complete, final brief — stream to the browser in real-time via Server-Sent Events over Redis.
RunnableParallel · SSE · RedisReal architectural decisions — not toy demos.
LangChain's RunnableParallel runs all 7 analysis chains at once via Celery — cutting sequential analysis from 3+ minutes down to the time of the single slowest chain (~20-30s).
Documents are embedded with BAAI/bge-small-en-v1.5 (384-dim) and indexed in Pinecone Serverless. At query time, cosine similarity results are re-ranked by keyword overlap — hybrid scoring for better grounding.
Pipeline events are pushed to Redis by the Celery worker and streamed to the browser via Server-Sent Events — users see every step live as it happens, not a loading spinner.
Every agent output is validated against a Pydantic v2 schema via LangChain's JsonOutputParser. Failed parse triggers an automatic JSON-only retry. No unstructured text reaches the DB.
Market Analysis, Competitive, and Roadmap agents use chat_model.bind_tools() — the LLM calls quantitative and qualitative tools mid-chain. Tool results are injected back before the final JSON response.
After all 7 chains complete, a final aggregation chain synthesizes everything into a coherent strategic brief. Agent outputs are slimmed to ~400 tokens before aggregation to stay within context limits — a deliberate token-efficiency choice.
The HuggingFace transformer (BAAI/bge-small-en-v1.5) and Pinecone index client are cached as module-level globals in the Celery worker process. They load once on first task, then are reused across all subsequent pipeline runs — eliminating a ~17s model-load penalty and 14 wasted Pinecone HTTP calls per pipeline.
A full-stack async system — from auth to vector search to real-time UI.
The complete set of technologies used — from LLM orchestration to deployment infrastructure.
Upload your knowledge base, set your objective, and get a full AI-powered strategic brief — streamed live in under 30 seconds.