AI model
Gemini and Claude were ported from the test harness but not yet live-tested end to end here -- verify with a throwaway submission before relying on them.
System prompt
The shared prompt for the distribution check, used by every state that does not have its own (set per state below). Placeholders: {{species}}, {{q2}}-{{q6}} (surveyor's own answers, filled in automatically).
State settings
Everything that differs by state lives here: the health questions the field team is asked, the regional breed notes, and the state's own AI prompts. It saves one state at a time, and every change is kept in the history with who made it and when.
AI check on the phone
With an animals-per-household number set (below, the usual case), the phone does not wait for the AI after each animal: every animal's check is shown together once the household is complete. Otherwise, when this is on and the phone has signal, Submit is followed by a screen with the AI's verdict on whether the animal is good enough to buy, its comments and its recommendations, next to the field worker's own answers (it usually takes 10 to 20 seconds). The entry, with the field worker's answers, is already saved before the AI is asked, so the AI never changes what they recorded. The screen then asks whether they agree with the AI, and if the AI's verdict differs from their first call (good enough to buy against Accepted, not good enough against Rejected) whether they change their decision or keep it. Their final call is logged next to the first one (the sheet's Final Decision, Agrees With AI Check and Decision Changed After AI columns); the first call is never overwritten. They can also save the check as a picture on the phone, or share it on WhatsApp and the like. With no signal the entry is kept on the phone, and as soon as the phone is online again it is sent and a note at the top of the screen says the check is ready to read. When this is off, nothing changes: the check is made in the background and is in My Submissions afterwards. This is for the distribution check; West Bengal's procurement photos are reviewed in batches during the day.
Poultry check on the phone
When it is on, the phone shows a "What are you checking?" box under "Where are you working?", with goats and pigs already chosen; a worker picks Poultry to do the poultry check, and one who ignores the box gets the usual form. A state where it is off never shows the box.
Poultry health check on the phone
Only matters where the poultry check is on. Batches checked the short way show on the dashboard's Poultry tab in their own card; the twelve per-part figures count only the batches that had the twelve checklists.
Animals per household
How many animals the surveyor accepts for one household visit in this state (rejected animals do not count). With the AI check on, the phone does not make the field worker wait for it after each animal: the animals are sent while the next one is examined, and every animal's AI check is shown together once this many have been accepted, each with its "do you agree?". Without a signal, or with 0, the check is made animal by animal as before. Rajasthan starts at 2 goats; other states at 3 goats and 2 pigs.
Health questions
What the field team is asked about an animal's health in this state. Change the wording, switch a question off, add your own, or put them in a different order. The AI check and the dashboard follow this list.
Regional breed notes
Added after whichever prompt is used (the shared one or this state's own below). Use it for things like a locally common breed's typical size or coat, so the AI doesn't misread a normal regional variation as an anomaly.
Distribution check: this state's own prompt
Replaces the shared System prompt for this state only. Same placeholders. Empty = this state uses the shared prompt. Changes apply to checks made after you save, not to ones already done.
Procurement review: this state's own prompt
Replaces the shared Procurement review prompt for this state only. Same placeholders. Empty = this state uses the shared one.
Procurement AI review
Goats bought by didis (West Bengal) are reviewed from their two photos, not while the phone waits, but in a batch three times a day: about 2:00 PM, 6:30 PM and 3:00 AM. The results appear on the dashboard's Procurement tab. It uses the AI provider and model chosen at the top of this page.
Placeholders: {{species}}, {{age_months}}. The UPCM's own ratings are not shown to the AI by default, so its opinion stays independent; to show them, add {{upcm_skin}}, {{upcm_back}} (goats) or {{upcm_health}}, {{upcm_legs}} (pigs). Keep the JSON reply format at the end as it is: the dashboard reads those exact fields.
Save changes first: a review uses the saved prompt. It starts in the background within a minute or two; reload this page afterwards to see the result above. Animals already reviewed keep their verdict.
Ear tag photo reading
Prompt used when a surveyor takes a photo of the physical ear tag instead of typing it (see the field app's Ear Tag step). A dedicated, focused prompt -- reading printed characters is a different task from the health assessment above.
Field-app dropdowns
The health questions (formerly Q2 to Q6 here) are now edited per state, under State settings above.
Saves the settings above this bar (AI model, shared prompts, dropdowns). State settings have their own Save button.
Retry failed AI checks
Re-runs the AI check for every submission still marked failed (bad/missing key, rate limit, provider outage, etc.) using its already-uploaded photos. Only rewrites AI columns in place -- never touches the surveyor's own answers or creates a new row. Processes up to 20 per click; click again if more remain.
API keys & secrets
Write-only -- existing values are never shown here, only whether one is set. Paste a new value and click Update to rotate it.