gptoss-demo

gptoss-demo — The Open-Source Copilot for HR, Compliance, and Life Insurance Guidance.”

Comment

Inspiration

Enterprises struggle with fragmented knowledge sources (HR policies, compliance docs, regulations). Employees often waste time or make mistakes because AI tools are either closed-source, cloud-only, or lack governance. We wanted to show that open-source models like gpt-oss-20B can power enterprise-grade copilots that are private, compliant, and extendable.

What it does

gptoss-demo is a minimal copilot that: Answers HR, Occupational Health & Safety (OHS), and Human Rights questions. Switches automatically between general HR mode and Life Insurance–grounded mode. Uses a Hugging Face–hosted text file (Life Insurance Basics.txt) to provide referenced, cited answers. Runs on open-source GPT-OSS-20B via Hugging Face Router — no vendor lock-in. Provides a clean Next.js UI with sample prompts, auto-detected domain badges, and citation rendering.

How we built it

Frontend: Next.js (App Router) + React + ReactMarkdown for clean, markdown-rendered answers. Backend: Next.js API route (route.ts) that: Detects domain (HR vs Life Insurance). Fetches and chunks text from Hugging Face dataset. Builds context-aware prompts with citations. Calls Hugging Face Router (openai/gpt-oss-20b) for completions. Retries gracefully if responses are cut off. Deployment: Vercel for the web app + Hugging Face for dataset hosting.

Challenges we ran into

Handling long context and token cut-offs with 20B — solved with retry logic and excerpt clipping. Making responses cite excerpts correctly without leaking internal reasoning. Keeping responses under Vercel’s serverless timeouts (15s). Debugging Hugging Face provider suffixes (:fireworks-ai vs router auto-pick).

Accomplishments that we're proud of

Built a working multi-domain copilot demo in under hackathon time. Integrated open-source 20B model into a Next.js app without OpenAI’s closed stack. Automatic domain switching (HR ↔ Life Insurance) with context-based grounding. Clear UI with sample prompts and reference citations for transparency.

What we learned

Hugging Face Router sometimes returns only “reasoning_content” — you need retry logic and token caps to get stable answers. Chunking + scoring + clipping are essential for grounding large models with external files. Users/judges really appreciate transparency (references, disclaimers, mode badges).

What's next for gptoss-demo

Add multi-file ingestion (HR handbooks, safety manuals, contracts). Use a vector database (e.g., Supabase/pgvector) instead of naive chunk-scoring. Extend to Finance & Legal copilots under the same FluxCore architecture. Fine-tune GPT-OSS models with Canadian labor law corpora for SME compliance. Add enterprise features: SSO, audit logs, dashboards.

Built With

Share this project:

Updates