Finora

Finance teams drown in PDFs, statements, and P&L exports. They want a copilot they can trust with numbers — not a chatbot that hallucinates.

• ▸ Agentic RAG with LangGraph nodes for parse → retrieve → reason → cite.
• ▸ ChromaDB embeddings per workspace; deterministic chunking on financial tables.
• ▸ Output JSON schema enforced via function-calling — every number is grounded in a cited chunk.
• ▸ Guardrails reject answers when retrieval confidence drops below a threshold.

[PDF/CSV] → ingest worker → ChromaDB (per-tenant)
                          ↓
           LangGraph: classify → retrieve → reason → cite
                          ↓
          structured JSON  →  React dashboard + chat

• ▸ Reduced hallucination rate to <2% on internal eval set.
• ▸ Cut analyst document-review time by ~70% in pilot tests.

• ▸ Retrieval quality beats model size for finance — chunking strategy was the biggest unlock.
• ▸ Hard schemas + 'I don't know' as a first-class output build user trust faster than higher accuracy.