Each card is one decision from the build — the path we dropped, the one we kept, and the reason it changed.
Raise max_tokens and let it reason freely
reasoning_effort: low — there is no “off”
Reasoning is mandatory on 105B and counts against the output cap. Left free, it used the whole token budget, so the answer came back empty with finish reason “length”.
Assume the shared Sarvam pipeline would match the input language
Detect the language on paste, then tell the model to reply in it
The prompts were written in English and nothing named a target language, so the council answered in English. Language detection now flows from the paste into every call.
Read the results response as plain JSON text
Check for the gzip marker and unzip it by hand
Doc AI returns results gzipped but without a Content-Encoding header, so nothing unzips them automatically. We now sniff the first two bytes (1f 8b) and gunzip.
output_format: “markdown”
output_format: “md”
The API rejected “markdown” outright. It only accepts “html” or “md”, which the docs don’t spell out.
Claim support because one Hindi run looked fine
Run the full pipeline in Kannada, Hindi and Tamil and record the numbers
One good run isn’t evidence. We now record tokens, cost, how much of the output is in the native script, and any truncation — so the support claim is backed by data.
Take whatever the model returns on the first try
Detect an empty, truncated reply and re-roll once
In one full run, a single persona hit the token cap and returned nothing. A guard now retries once before giving up, so a run doesn’t ship with a missing voice.
Send the clip as its real type, audio/webm;codecs=opus
Send it as application/octet-stream
Saarika whitelists by the declared file type and rejects webm;codecs=opus — the exact thing a browser’s MediaRecorder produces — even though it decodes the audio fine. Sent as octet-stream, it sniffs the real container, and it works the same across Chrome, Firefox and Safari.