Multi-agent RAG · Groq · Observable backend
Analyze business documents with intelligent agents
Upload sales, finance, or ops files, then ask in plain English. Grounded answers, cited retrieval, and a routing layer built for portfolio-grade demos.
API base: /agentflow-api. Production: use NEXT_PUBLIC_API_URL=/agentflow-api on Vercel (proxied to Render to avoid CORS). See repo README.
What you’re looking at
Three steps that map to the architecture: ingest → retrieve → answer. No jargon required to try the demo; the labels in the app mirror how the backend works.
1 · Bring your file
Upload or use sample data. The API parses and stores the document against your session.
2 · Ask a real question
Plain English. The orchestrator may route to RAG (search your chunks), Q&A, summary, or verify.
3 · Read the answer
Streaming replies. When retrieval runs, expand “sources” to see which text chunks were used.
See it in action
Preview of the workspace below: upload, chat, and answers grounded in your documents. Use Open workspace above to try it live.
What this preview shows
The same flow you get in the live app
- 01Work with business files (CSV, Excel, PDF, text).
- 02Ask questions in natural language; the backend retrieves relevant chunks, then replies.
- 03Replies can include source passages so you see what the model used.
Shipped on purpose: a silent preview of the real UI. Replacing it with a narrated video is optional; steps are in the project repository for you when you want them.
Production-style architecture
Next.js frontend, Express orchestration, embeddings + vector search, file-backed persistence, telemetry and eval hooks, documented for interviews and code review.
Orchestration
Routes queries to ingest, RAG, Q&A, verifier, or summarizer agents.
API layer
REST + SSE streaming; workspace headers for scoped sessions.
Retrieval
Chunking, embeddings, cosine similarity; swap for pgvector later.
Persistence
JSON-backed sessions and vectors; configurable DATA_DIR for hosts.
Observability
Telemetry routes, Prometheus metrics, eval run history.
Docs
Full diagrams and flows in the architecture guide.
Built for real workflows
Document analysis
Revenue sheets, briefs, and ops updates: summaries, metrics, and grounded Q&A.
RAG retrieval
Semantic search surfaces the right evidence; responses can cite sources and scores.
Multi-agent
Specialized agents for analysis, verification, and summarization, routed for you.
How it works
End-to-end path from upload to grounded answer; same mental model you’d use in a system design interview.
Upload
CSV, Excel, PDF, or text
Index
Chunk + embed for semantic search
Route
Orchestrator picks the best agent
Answer
Streamed response + optional sources
Agent roster
Each role is explicit in the UI and in routing, easy to extend with new tools or policies.
Retrieval-augmented answers
Direct Q&A with context
Claims and consistency
Briefs and key points