Umbrella repository for CropWizard work under AIFARMS: the public website, the agentic tools built for the CropWizard assistant, and the operational record of the Supabase/pgvector corpus migration.
The public CropWizard site, served via GitHub Pages at
https://aifarms.github.io/cropwizard-project/. A static Bulma-based site presenting the
CropWizard benchmark for multimodal expert-level reasoning in agricultural consultative
interactions — leaderboards (leaderboards/), category and LLM-judge case studies, and
figures.
Note: GitHub Pages is currently configured to publish from the
docs/folder onmain. This folder was renamed fromdocs/towebsite/, so the Pages deployment settings must be updated (legacy Pages only supports/or/docs; use a GitHub Actions Pages workflow or repoint the source) or the live site will stop updating.
The tools built for CropWizard — each subfolder is one tool, deployed independently (Beam Cloud endpoints and/or n8n workflows):
| Tool | Purpose |
|---|---|
csv_geo_agent/ |
Geospatial agent over field/urban-agriculture CSV + GeoJSON data |
google_web_search/ |
Google web search endpoint |
MRTN_tool/ |
Maximum Return To Nitrogen (MRTN) rate calculator |
pest_detection/ |
Pest detection from images |
weather_agent/ |
Weather + soil agent (plain, RAG, and GCP variants) |
Copy tools/env.example to .env and fill in credentials before running any tool.
Note: this folder is intended to track every tool we build and deploy for CropWizard, but it is out of date and needs a cleanup pass (consistent structure, per-tool READMEs, deployment docs). Kept as-is for now.
The complete record of re-ingesting the cropwizard-1.5 corpus (~684k documents) from
the original NCSA-hosted Supabase + Qdrant into a fresh Supabase project with pgvector,
re-embedded with Qwen3-Embedding-8B (June 23 – July 8, 2026). Contains the migration
scripts, AWS infra definitions, run data, and a detailed write-up of the methodology,
failure taxonomy, and final state — see supabase-reingest/README.md.
The raw run logs in supabase-reingest/logs/ are tracked too — they were scanned for
credentials and contain only submitted URLs/filenames and progress counters.
See LICENSE.