Document Triage Pipeline

Overview

An offline batch pipeline that classifies and routes large document sets without an interactive latency budget.

This page shows the structure a blueprint detail page uses. The problem statement and component list below are real design context; the deployment and usage sections are placeholders.

Problem addressed

Bulk document classification is throughput work, but it is often built on interactive serving infrastructure, which wastes most of the available accelerator capacity.

Architecture

Component overview

  [1] Batch scheduler
  [2] Classification prompts
  [3] Structured output schema
  [4] Checkpoint store
  [5] Inference backend

A concrete architecture diagram will replace this sketch when the blueprint is published.

Key components

Batch scheduler — Feeds documents in size-bucketed batches to maximise throughput.
Classification prompts — Versioned prompts with a fixed evaluation set.
Structured output schema — Constrains each result to a parseable record.
Checkpoint store — Makes a long run restartable after interruption.
Inference backend — An llm-inference deployment configured for throughput.

Prerequisites

An inference endpoint. On MareNostrum 5, start one with llm-inference api start — see the quickstart.
The base URL of that endpoint, http://<host>:<port>, as printed by api start.
A model ID listed by llm-inference model list with SUPPORTED=yes.

Deployment instructions

Not yet available

This is demonstration content. No deployable artifacts exist for this blueprint, so no deployment commands are given here — an invented command would be worse than an acknowledged gap.

When the blueprint is published, this section will contain the repository location and the steps to deploy it.

Configuration

The configuration surface will be documented here. It is expected to cover at least:

Setting	Purpose
`<ENDPOINT_BASE_URL>`	Base URL of the inference endpoint, e.g. `http://nid001:45123`.
`<MODEL_ID>`	Identifier of the model to call.

Usage

Draft

Worked usage examples will be added with the published blueprint.

Security considerations

These apply to this blueprint’s design and are not placeholders — they hold regardless of implementation:

Model output is untrusted input. Validate every tool call independently before acting on it. See Function calling.
The inference endpoint has no authentication. api start exposes a plain HTTP port inside the MN5 network; anything reachable from there can use your allocation.
Enforce authorization in your own code, on the end user’s identity — not on what the model requests.
Treat prompt content as data subject to your retention and access rules; it is written to the shared filesystem.

The full guidance is in Security notes.

Observability

Per-request latency and success or failure are recorded for every request; where those land and how to inspect them is described in Logs and monitoring. Application-level outcome metrics specific to this blueprint will be listed here on publication.

Limitations

Warning

This page is demonstration content. It has no implementation behind it, no measured performance characteristics and no validated deployment path. Do not treat the architecture sketch above as a tested design.

Next steps

Read the blueprint catalog for other blueprints.
Get an inference endpoint running: llm-inference quickstart.
Understand the serving concepts involved: LLM Inference Handbook.