together-batch-inference

name	together-batch-inference
description	High-volume, asynchronous offline inference at up to 50% lower cost via Together AI's Batch API. Prepare JSONL inputs, upload files, create jobs, poll status, and download outputs. Reach for it whenever the user needs non-interactive bulk inference rather than real-time chat or evaluation jobs.

Together Batch Inference

Overview

Use Together AI's Batch API for large offline workloads where latency is not the primary concern.

Typical fits:

• bulk classification
• synthetic data generation
• dataset transformations
• large summarization or enrichment jobs
• low-cost asynchronous inference

When This Skill Wins

• The user has many independent requests to run
• A JSONL request file is acceptable
• Turnaround time can be minutes or hours instead of seconds
• Lower cost matters more than immediate interactivity

Hand Off To Another Skill

• Use together-chat-completions for real-time requests or tool-calling apps
• Use together-evaluations for managed LLM-as-a-judge workflows
• Use together-embeddings for retrieval-specific vector generation

Quick Routing

• End-to-end batch workflow
  • Start with scripts/batch_workflow.py or scripts/batch_workflow.ts
• Request format, status model, and result downloads
  • Read references/api-reference.md
• Operational guidance and batch sizing
  • Read references/api-reference.md

Workflow

1. Build a JSONL file where each line contains custom_id and body.
2. Upload the file with purpose="batch-api".
3. Create the batch with input_file_id=... and the target endpoint.
4. Poll until the job is terminal.
5. Download output and error files, then reconcile by custom_id.

High-Signal Rules

• Python scripts require the Together v2 SDK (together>=2.0.0). If the user is on an older version, they must upgrade first: uv pip install --upgrade "together>=2.0.0".
• Use input_file_id, not legacy file parameters.
• Keep custom_id stable and meaningful so result reconciliation is easy.
• Batch is for independent requests. If the workload depends on shared conversation state, it is probably the wrong tool.
• Always inspect the error file in addition to the success output.
• client.batches.create() returns a wrapper; access the batch object via response.job (e.g., response.job.id). client.batches.retrieve() returns the batch object directly.
• For classification or labeling workloads, set max_tokens low (e.g., 4), use temperature: 0, and constrain the system prompt to return only the label. This minimizes output tokens and cost.
• Small batches (under 1K requests) typically complete in minutes. The 24-hour completion window is a maximum, not typical.

Resource Map

• API reference and operational guidance: references/api-reference.md
• Python workflow: scripts/batch_workflow.py
• TypeScript workflow: scripts/batch_workflow.ts

Official Docs

• Batch Inference
• Batch API