CrediSense AI: Credit Policy RAG Engine

Inspiration

Commercial credit policy manuals, banking regulations, and loan underwriting guidelines are dense documents spanning hundreds of pages. Credit officers and loan underwriters spend countless hours manually cross-referencing covenant ratios (such as DSCR, Current Ratio, and LTV caps), regulatory thresholds (such as LEI and CRILC reporting), and pre-disbursement compliance checklists.

A single missed covenant or miscalculated limit can result in severe audit penalties or loan defaults. While general-purpose LLMs are powerful, they frequently hallucinate numbers and cannot be trusted in financial risk assessments. We were inspired to build CrediSense AI—a zero-hallucination, 100% grounded credit policy RAG copilot that provides instant, page-exact audit citations from the bank's active policy manual.

What it does

CrediSense AI transforms dense institutional loan policies into an interactive, real-time credit underwriting copilot:

• Single Active Policy Document Workflow: Upload or update any loan policy PDF (e.g., RBIC Loan Compliance Policy, SBA Underwriting Guide). The system cleanly wipes previous vectors and embeds the new manual instantly.
• Real-Time Token Streaming with Citations: Delivers answers token-by-token in under 2 seconds, backed by exact page numbers, chunk IDs, and cosine similarity scores.
• Strict Ground Truth & Zero Hallucination: If a policy criterion or ratio isn't present in the active document, the AI explicitly states that it is not covered rather than making assumptions.
• Interactive Vector Explorer: Allows risk officers to inspect semantic embeddings, similarity distributions, and raw chunk contexts directly inside ChromaDB.
• Audit Report Generation: Exports comprehensive Markdown audit reports containing dialogue turns, grounded answers, and verified source citations with a single click.

How we built it

• Embeddings & Vector Store: Built with nomic-ai/nomic-embed-text-v1.5 generating 768-dimensional asymmetric vector embeddings (search_document vs search_query prefixes) stored in a persistent ChromaDB collection with cosine similarity indexing.
• LLM Reasoning & Generation: Integrated Groq-accelerated open-weights inference (openai/gpt-oss-120b and fallback models) via streaming Server-Sent Events (SSE).
• Backend Architecture: High-performance FastAPI backend with automated PDF extraction, semantic text chunking, and multi-turn dialogue memory management.
• Frontend Experience: Modern, responsive UI built with Vanilla JavaScript and CSS featuring dark-mode glassmorphism, live status indicators, preset loan scenarios, and collapsible citation cards.

Challenges we ran into

1. Eliminating Financial Hallucinations: Ensuring the LLM strictly adhered to the retrieved context without inventing realistic-sounding interest rates or covenants required rigorous system prompt engineering and similarity threshold filtering.
2. Asymmetric Embedding Precision: Tuning the 768-dim Nomic embedding pipeline to distinguish nuanced banking terminology (e.g., distinguishing DSCR stressed floor of $1.25\times$ vs standard $1.50\times$) using task-specific prefixing.
3. Seamless Single-Document Re-indexing: Implementing a race-condition-free document replacement pipeline that completely cleans up the vector database and re-embeds new policies on-the-fly without server restarts.

Accomplishments that we're proud of

• Sub-2-Second End-to-End Latency: Achieving real-time semantic retrieval and token streaming.
• 100% Grounded Traceability: Every response includes transparent, verifiable source cards with exact page numbers and chunk snippets.
• Frictionless Document Swapping: Underwriters can drop any new credit policy PDF and immediately begin testing without configuring databases or pipelines.

What we learned

• The critical importance of task-specific prefixing in asymmetric vector embedding models for high-precision retrieval.
• How to structure resilient Server-Sent Events (SSE) streaming with multi-turn memory buffers.
• Practical insights into commercial credit risk standards, including Large Exposure Frameworks, SMA-0/1/2 delinquency classifications, and pre-disbursement compliance workflows.

What's next for CrediSense AI

• Automated Borrower Financial Ingestion: Uploading borrower balance sheets and P&L statements alongside the policy to automatically calculate DSCR/LTV and flag policy breaches.
• Side-by-Side PDF Citation Viewer: Embedding an in-app PDF viewer that highlights the exact cited text directly on the visual page.
• Multi-Policy Circular Comparison: Comparing internal bank credit policies against newly issued central bank/regulatory circulars to detect policy gaps.

Built With

• api
• artificial
• chromadb
• embeddings
• events
• fastapi
• fintech
• groq
• html5
• intelligence
• javascript
• llm
• nomic
• prompt
• pydantic
• pypdf
• python
• rag
• rest
• search
• sentence
• server-sent
• transformers
• uvicorn
• vector