How do you build summarization with AI agents?

Key facts

Pattern	Split → summarise sections → combine → verify coverage
Runtime	Chat completions with streaming and JSON mode (live)
Long context	plugsky-longctx for combining many partial summaries
Citations	Section-level source references in structured output
Models	30+ models behind one endpoint, route per pass
Verification	Coverage checks against extracted entities and figures
Deployment	Region-locked planes, VPC, on-prem and air-gapped options
Free tier	plugsky-micro and plugsky-lite on the free plan, no card required

TL;DR

• Summarise in passes for long inputs; single-shot prompts lose the middle.
• Keep citations per section so readers can verify any claim.
• Extract figures and entities first, then check the summary covers them.
• Use JSON mode for summaries with sections, bullets and sources.
• Evaluate coverage and factual fidelity, not just readability.

How it works, step by step

1. Estimate input size in tokens and decide between single-pass and multi-pass summarisation.
2. Split long inputs along natural boundaries — sections, speakers or document structure — keeping references.
3. Summarise each part with an instruction to preserve figures, dates, names and decisions.
4. Combine partial summaries with a long-context model, requiring citations to the original parts.
5. Run a coverage check: extract key entities and figures from the source and confirm they appear.
6. Add a verification pass that flags unsupported or missing items before publication.
7. Evaluate coverage, factual fidelity and length targets on a labelled set of documents.

Try it yourself

Open the tokenizer →

Map, reduce and verify

Long inputs break single-shot summarisation: models compress beginnings and endings and lose the middle. The fix is structural — summarise in parts, then combine. Each part keeps its source reference, and the combination step works from partial summaries rather than the full text, which keeps the final prompt manageable.

Plugsky's chat completions endpoint is live with streaming and JSON mode, so summaries can be structured objects with sections, bullets and sources. When the agent needs to fetch related material — a referenced appendix, a linked thread — function calling handles it in the same loop.

Chunking and long-context trade-offs

There are two valid strategies. Multi-pass map-reduce handles very long inputs reliably and scales with document size. Single-pass with a long-context model is simpler and better at preserving cross-section reasoning when the input fits comfortably. Choose per document type rather than committing to one approach.

• Preserve structure: split on sections, headings or speaker turns, not fixed windows.
• Protect facts: instruct each pass to retain figures, dates and named entities verbatim.
• Reference everything: every bullet in a partial summary cites its section.
• Mind the budget: check token counts before prompting and trim retrieval rather than truncating mid-sentence.

Evaluating summaries

Readability is the easy part; fidelity is what matters. Evaluate coverage — are the key entities, figures and decisions present — and factual accuracy — does every claim match the source. Build a labelled set of documents with the facts that must survive, then score each summary against it.

Track length targets too, since summaries that balloon or truncate fail their purpose. Route section summaries to small, fast models and the combination pass to a long-context model. Store the structured summary with its citations so reviewers and downstream systems can verify claims without reopening the source. For confidential documents, deploy in a region-locked plane or on-prem with Plugsky and keep retention aligned to policy.

Honest comparison

Concern	Agentic multi-pass	Single-shot prompt	Manual summarising
Long inputs	Handles full documents reliably	Loses middle sections	Depends on reader
Citations	Section-level references retained	Usually absent	Notes
Fact preservation	Explicit figures and entities	Inconsistent	Human judgement
Verification	Coverage checks before release	None	Review
Cost shape	More calls, smaller prompts	One large call	Human time

Frequently asked questions

When should I use multi-pass summarisation?

When the input exceeds what a single prompt handles well or contains many distinct sections. Splitting on structure and combining partial summaries preserves coverage far better than one large prompt.

How do I stop summaries losing numbers?

Instruct each pass to retain figures, dates and named entities verbatim, then run a coverage check that compares extracted facts against the final summary and flags anything missing.

Should summaries include citations?

Yes. Section-level references let readers verify claims quickly and make the summary auditable, which matters for meetings, research and compliance work.

Which model handles the combine step?

A long-context model such as plugsky-longctx works well when many partial summaries must be weighed together. Section passes can run on smaller, faster models.

How do I evaluate quality?

Score coverage of must-keep facts, factual fidelity against the source and adherence to length targets, using a labelled document set rather than subjective reading.

Can the agent summarise meeting transcripts?

Yes. Split by speaker turn or topic block, summarise each, then combine into decisions, actions and open questions with references back to the transcript.

Can it run on-prem?

Yes. Plugsky supports region-locked cloud planes plus VPC, on-prem and air-gapped deployment, so confidential documents can stay inside your network.