Key facts
| Router model | plugsky-fusion escalates per question across tiers (live) |
| Embeddings | Live embeddings API; re-embed only changed documents by hash |
| Models | 30+ models; lookups on cheap tiers, synthesis escalated |
| Citations | JSON mode for claim-to-source mapping |
| Pricing | Flat monthly self-serve plans with no per-token charges on self-serve |
| Free tier | plugsky-micro and plugsky-lite on the free plan, no card required |
| Long context | plugsky-longctx for synthesis spanning many passages |
| Roadmap | Files and batch endpoints are coming soon; keep ingestion in your pipeline |
TL;DR
- Index incrementally; unchanged documents should never be re-embedded.
- Retrieve tighter: fewer passages, higher relevance thresholds.
- Cheap tiers answer lookups; synthesis escalates deliberately.
- Citations keep quality measurable at any tier.
- Cost per answered question is the metric that matters.
How it works, step by step
- Hash documents and re-embed only new, changed or restored content.
- Tune chunk size and top-k against answer quality, not index size alone.
- Classify questions into lookup, comparison and synthesis types.
- Route lookup questions to cheap tiers with tight context and citations.
- Escalate synthesis to stronger tiers or plugsky-longctx when breadth is genuinely required.
- Cache question embeddings and answers where content freshness allows.
- Track cost per answered question and citation accuracy by question type.
Try it yourself
Open the RAG cost calculator →
Knowledge retrieval dominates
Before a knowledge assistant answers anything, someone indexed the corpus, and that cost recurs as documents change. Incremental indexing by content hash keeps maintenance proportional to change rather than to corpus size, and it removes the most common source of runaway spend: nightly full re-embeds.
At query time, retrieval decides both quality and cost. Every extra passage is paid for in context on that question and every similar question after it. Tune top-k and relevance thresholds so the prompt contains evidence, not archive. Better retrieval is the only optimisation that improves quality and reduces cost simultaneously.
Answer quality versus model size
Lookup questions — where a single passage contains the answer — are solved by retrieval plus a small model. Synthesis questions reward stronger models because they require weighing multiple sources and resolving conflicts. Treating both the same either overpays for lookups or under-delivers on synthesis.
- Ask for citations in a structured form so quality is checkable at any tier.
- Refuse politely when evidence is missing instead of guessing.
- Use plugsky-longctx when synthesis needs breadth, not as a default.
- Keep a golden question set per topic to catch regressions after content changes.
Freshness, caching and cost accounting
Knowledge moves. Expire superseded documents, version the index, and invalidate cached answers when their sources change. Where content is stable, cache answers with a time-to-live so repeat questions cost almost nothing while freshness stays controlled.
Report cost per answered question alongside citation accuracy, answer usefulness and refusal quality. A cheap answer that cites the wrong policy is the most expensive outcome in the system. Start on the free plan with plugsky-micro and plugsky-lite for the pipeline and evaluation, then use the 14-day full-access trial to test stronger and long-context models on synthesis questions; the live pricing page lists plans.
Honest comparison
| Knowledge assistant choice | Routed assistant on Plugsky | Strong model for every answer | Cheap model for every answer |
|---|---|---|---|
| Lookup questions | Cheap tier, tight context | Frontier price for facts | Adequate |
| Synthesis | Escalated or long-context model | Native strength | Shallow and incomplete |
| Index upkeep | Incremental by hash | Often full rebuilds | Often full rebuilds |
| Repeat questions | Cached answers with expiry | Paid again | Paid again |
| Quality evidence | Citation accuracy by tier | Rarely measured | Rarely measured |
Frequently asked questions
What costs more: indexing or answering?
Indexing dominates for large, fast-changing corpora, while answering dominates for high query volume on stable content. Measure both, then fix the larger one first.
How do I cut indexing cost?
Embed incrementally: hash documents and re-embed only new or changed content. Nightly full rebuilds are the most common avoidable expense in knowledge systems.
Should all answers use a strong model?
No. Single-passage lookups work well on cheap tiers with citations. Synthesis across documents benefits from stronger or long-context models, so route by question type.
How many passages should retrieval return?
Tune top-k and thresholds against answer quality. Sending more passages raises cost on every question and adds noise that can reduce accuracy.
How do I keep answers current?
Version your index, expire superseded documents, and invalidate cached answers when their sources change. Citations make staleness visible to users.
Can I cache answers?
Yes, with a time-to-live tied to content change frequency. Cache question embeddings always, since repeated questions are common.
How do I evaluate knowledge quality?
Keep a golden question set per topic with expected sources, and track citation accuracy, correctness and refusal quality after every content or model change.
Can I start for free?
Yes. plugsky-micro and plugsky-lite are on the free plan with no card, and the 14-day full-access trial covers stronger and long-context models.