Key facts
| Long-context tiers | plugsky-longctx and plugsky-kimi for many-page inputs |
| Fast long-context | plugsky-nano for longer prompts where speed still matters |
| Deep analysis | plugsky-max for the hardest multi-document reasoning |
| Retrieval companion | pair long-context chat with plugsky-embed for chunked retrieval |
| Context class | 128K-class windows on current models; live limits are published per model |
| Capabilities | Streaming, function calling and JSON mode on the OpenAI-compatible API |
| Free plan | plugsky-micro and plugsky-lite, no card required |
| Product status | Live |
TL;DR
- Test plugsky-longctx and plugsky-kimi first for many-page inputs.
- Use plugsky-nano when long prompts still need to stay fast.
- Reserve plugsky-max for multi-document reasoning and synthesis.
- Above a few documents, chunk-and-retrieve with plugsky-embed usually wins.
- Check the live model card before hard-coding any window size.
How it works, step by step
- Measure the real input size: pages, tokens and how often inputs exceed a single window.
- Test plugsky-longctx and plugsky-kimi on a sample of your longest documents.
- Compare a single long-context call against chunk-and-retrieve with plugsky-embed.
- Check the live model card for the window, output limit and capability flags.
- Add a summarise-then-answer stage for corpora that exceed the window.
- Track answer accuracy and cost per document, then route by input length.
Try it yourself
Open the context window comparison →
Long context or retrieval?
Long-context models are the simplest option: send the whole document and ask your question. That works well for one or two documents, legal clauses, contracts and reports where the model must see everything at once. It breaks down when the corpus keeps growing — every request pays for the entire input, answers drift, and you cannot cite a source you never indexed.
Chunked retrieval with plugsky-embed inverts the trade-off: index once, retrieve only the relevant passages, and answer with a smaller context. The practical rule is to prototype with long context and move to retrieval when document count, update frequency or citation requirements increase.
Which long-context tiers to test
Start with plugsky-longctx and plugsky-kimi, both aimed at long-input work. Add plugsky-nano when the same workload must also stay responsive, and plugsky-max when the task is genuinely analytical — comparing clauses, reconciling figures or synthesising several sources into one answer.
Windows and output limits change as models are updated, so read the live model card on /models rather than trusting a number from a blog post, including this one.
Patterns for very large corpora
Once inputs exceed what any single call should carry, use structure instead of a bigger prompt:
- Map and reduce: summarise each chunk independently, then summarise the summaries.
- Hierarchical retrieval: retrieve passages, expand to their sections, then answer.
- Citations first: require the model to quote the passage supporting each claim.
- Stable chunk IDs: keep document and section metadata so answers can be traced back.
These patterns keep cost predictable and answers auditable, and they work with the RAG API approach rather than against it.
Honest comparison
| Input pattern | plugsky-nano (fast) | plugsky-longctx, kimi | plugsky-max (top tier) |
|---|---|---|---|
| One long document | Good when speed matters | Best fit | Works, higher cost |
| Several related documents | Limited | Strong | Best fit for synthesis |
| Corpus that keeps growing | Use retrieval instead | Use retrieval instead | Use retrieval instead |
| Latency-sensitive summaries | Best fit | Good | Not ideal |
| High-stakes analysis | Not recommended | Good | Best fit |
Frequently asked questions
Which Plugsky model is best for long documents?
Start with plugsky-longctx and plugsky-kimi for many-page inputs, use plugsky-nano when latency matters, and plugsky-max for the hardest multi-document analysis. Compare them on your own documents before choosing.
Should I use long context or RAG?
Prototype with long context for one or two documents. When the corpus grows, updates frequently or needs citations, switch to chunk-and-retrieve with plugsky-embed and a chat model for synthesis.
What is the context window of these models?
Current models are 128K-class, but the exact window and output limit are published per model on the catalogue. Use the live model card as the source of truth.
Does long context get expensive?
It can, because every request carries the full input. Self-serve plans are flat monthly with fair-use usage, but the efficient pattern is still to send only what the task needs.
Can the model cite where an answer came from?
With retrieval, yes — you control the passages and can require quotes. With a single long-context call you can ask for page or section references, but accuracy depends on the model and document structure.
Can I try this on the free plan?
The free plan covers plugsky-micro and plugsky-lite with no card required, which is enough to prototype extraction and summary prompts. Use the 14-day full-access trial for the long-context paid tiers.
How is pricing structured?
Self-serve plans are flat monthly with fair-use usage rather than per-token billing. Check the live pricing page for current plans.