PARAS MADAN
All writing

Latest in AI

Sarvam Connect exports: ingest completed runs and recover missing windows

Build a reliable call-data export workflow around scoped agents, successful control files and explicit backfills instead of treating every uploaded archive as complete.

A scheduled export is only useful if the downstream report can tell a complete delivery from a missing one. Otherwise, an empty morning dashboard can look like a quiet day when it is actually a broken data pipeline.

Sarvam's November 18 Connect announcement adds scheduled delivery of voice-agent reports and recordings to SFTP or Amazon S3. The entry point is Monitor → Connect, where you add a connection and configure daily, weekly, monthly or historical one-time exports. Azure Blob and Google Cloud Storage remain forthcoming.

The useful implementation is not just a storage connection. It is a small ingestion workflow with a clear scope, a completion check and a recovery plan.

Start with one agent and one downstream purpose

The Connect guide says connection setup includes a reachability and sign-in test. Jobs can target selected agents or All agents; the latter includes new agents automatically at each run.

For an initial rollout, I would select one agent and the smallest dataset needed by its consumer. A daily operations report and an audio-quality review have different requirements. Do not copy recordings into an analytics destination simply because the option exists.

Write down who owns the destination, who reads it and what the receiving process will calculate.

Choose the right population for your report

The unified report includes all call attempts, while interactions includes connected calls only. Report CSVs and recording WAVs arrive in plain tar archives, not gzip. Debug calls are excluded.

This distinction changes the denominator of a business metric. Suppose you want to understand whether an outbound campaign is reaching people. Counting only conversations that connected cannot tell you how many attempts failed to connect.

Keep that calculation separate from conversation quality. A successful connection can still produce an unhelpful conversation, while an unanswered attempt cannot be scored as a conversation at all. Name the metric precisely before deciding which export to ingest.

Also test your archive reader against the actual format rather than relying on a filename pattern from another pipeline. Treat parsing failures as visible failures, not empty datasets.

Use completion metadata as the ingestion gate

Each agent folder has a control file. Process its data only when status equals success; a data type absent from that file was not produced. Individual files upload atomically, but file arrival alone is not the completion signal for the agent's export.

The receiving process should first read the control record, identify the described data and validate what it is about to load. Only then should it update the reporting tables.

Keep a separate ingestion ledger. Record the export location, agent, expected time window, observed completion and downstream processing result. These are suggested bookkeeping fields, not a claim about an additional Sarvam API.

Use the ledger to prevent duplicates. Re-reading yesterday's folder should not double yesterday's totals. Design the loader to recognize work it has already committed, and test that behavior before enabling a recurring job.

Separate scheduled time from completed delivery

All schedules use IST. Daily windows cover the preceding 24 hours, ending at the scheduled time. Jobs can start within 15 minutes of that time.

Do not make the dashboard refresh depend on a timer firing at the exact scheduled minute. Let ingestion wait for the completion condition, then publish a clearly marked reporting window.

For teams working across time zones, store explicit timestamps alongside the human-readable date. The label “yesterday” means different things to an operator in India and an analyst elsewhere. A report should explain which interval it covers without requiring either person to reconstruct the scheduler settings.

Plan backfills before the first missed delivery

Missed windows are not automatically backfilled. Historical one-time jobs cover inclusive past dates, up to 92 days, ending before today in IST.

Keep a list of expected windows and compare it with completed ingestion, rather than checking only the most recent successful folder. That is how you detect a gap hidden between two healthy deliveries.

A historical day-based export may overlap data already loaded from a scheduled window that ends at a different hour. Before recovery, map the missing interval and decide how overlapping records will be reconciled. Do not assume importing every recovered row is safe.

Run recovery into a staging area, compare counts and investigate discrepancies before updating the published report. Retain the original delivery files long enough to explain the reconciliation.

Make the first rollout easy to inspect

Use a short historical export to check archive extraction, column handling, completion gating and duplicate prevention. Then enable the recurring flow and verify that its reporting interval matches the intended business day.

Make missing delivery, failed parsing and genuinely empty results different states in monitoring. The point of owning the exported data is not merely having more files. It is being able to explain what the report includes, what it excludes and which period still needs recovery.

I'm Paras. I work at the intersection of AI, engineering and distribution.

I help AI teams explain integrations and production workflows through practical technical writing.

Book a 30-minute call.