Spaces:
Sleeping
Sleeping
Commit History
-
chore: remove sessions.db from tracking, add to gitignore
3ad88fd -
chore: add HuggingFace Space frontmatter to README
c885254 -
docs: rewrite README with badges, benchmark table, architecture diagram, and charts overview
e563233
-
docs: add README, requirements, and engineering problems log
README covers architecture, real benchmark results table, project structure, setup instructions, and metric definitions. PROBLEMS_FACED.md documents 15 real engineering problems encountered during development (Groq rate limits, LangGraph state serialization, Streamlit session resets, stream_mode format, etc.) with root causes and fixes.4f22cb2 -
feat: add animated Streamlit evaluation dashboard
Live Query panel: real-time pipeline progress bar, side-by-side agent responses, LLM-as-judge scoring on submission, 6-metric comparison cards. Benchmarks panel: 7 Plotly charts (5-metric radar, latency vs relevance scatter, hallucination gauges, category breakdown, win donut, trend line, latency distribution), 8 animated metric cards, response comparison tabs with mini metric progress bars per response. CSS: glassmorphism, blur-reveal number animation, shimmer pipeline steps, floating logo, live ping dot, gradient hero banners.8662564 -
feat: add LLM-as-judge evaluator and benchmark runner
LLM-as-judge (llama-3.1-8b-instant) scores responses on 5 metrics: relevance, hallucination, coherence, completeness, and depth. Resumable benchmark runner evaluates 50 queries across both agents, saves results incrementally to JSON to handle rate-limit interruptions. Results: multi-agent achieves 6x lower hallucination (4% vs 24%) with comparable relevance (0.85 vs 0.877) across 50 domain-spanning queries.b5122b0 -
feat: add core LangGraph multi-agent pipeline
5-node stateful pipeline: Planner -> Researcher -> Analyst -> Writer -> Memory. Each node uses Pydantic structured outputs for schema-enforced LLM responses. Includes FAISS vector store retrieval, Tavily web search, SQLiteDict session memory, and Python REPL tool for code execution tasks.be745f3