About ARIA
ARIA is an agentic retrieval‑augmented assistant for dryland and semi‑arid agriculture research. It searches the ICRISAT OAR corpus, extracts evidence, and produces grounded, citation‑ready answers for scientists, practitioners, and policy teams.
What makes it different
ARIA prioritizes research fidelity. It enforces section‑aware retrieval, tracks evidence at chunk level, and runs continuous evaluation against benchmark queries to keep the system honest.
Core Capabilities
Hybrid Retrieval
Combines semantic search, BM25 keyword search, and RRF fusion with diversity caps.
Workflow‑Aware Answers
Adapts the response style for factual, review, policy, comparative, or experimental requests.
Traceable Citations
Every answer is linked to chunk IDs and page numbers for verification.
Evaluation & Benchmarks
Runs a 90‑query suite across 9 workflows with detailed quality metrics.
Admin Analytics
Dashboards, metrics APIs, CSV export, and evaluation status tracking.
Security & Governance
CSRF, rate limits, audit logs, and safety review gates for high‑risk content.
System Workflow
1. Ingest & Index
Scrape OAR metadata, process PDFs, chunk by section, and embed for vector search.
2. Retrieve Evidence
Expand queries, run hybrid retrieval, apply section allowlists, and fuse results.
3. Synthesize
Generate responses tuned to the requested workflow with strict output contracts.
4. Evaluate
Score faithfulness, relevance, completeness, and citation support.
Who It Serves
Researchers
Rapid evidence synthesis for proposals, reviews, and hypothesis building.
Practitioners
Field‑level guidance based on vetted research, not anecdote.
Policy Teams
Clear briefs and recommendations grounded in ICRISAT publications.
Built With
Models & Infrastructure
Gemini 2.0 Flash, Qdrant or Vertex AI, LangGraph, and a PostgreSQL persistence layer.
Designed for traceability, not just fluency.
Governance
Safety checks for biosafety terms, audit logs, and admin‑controlled evaluation runs.