Use Cases + Implementation

How do you reduce the cost of enterprise search with model routing?

Enterprise search costs scale with corpus size and query volume, not with answer length. Embed incrementally so only changed documents pay, rerank selectively rather than on every result set, and reserve generated answers for queries that need synthesis. Keep keyword search in the mix so exact identifiers do not require expensive semantic work.

Key facts

Router modelplugsky-fusion escalates per query across tiers (live)
EmbeddingsLive embeddings API; re-embed only changed documents by hash
RerankingRerank selectively on high-value query classes
Models30+ models; answer generation only where synthesis is required
PricingFlat monthly self-serve plans with no per-token charges on self-serve
Free tierplugsky-micro and plugsky-lite on the free plan, no card required
GovernanceScoped keys, permission filters and audit logs
RoadmapBatch re-embedding is coming soon; run bounded refresh jobs today

TL;DR

  • Embed incrementally: only changed documents should cost anything.
  • Blend keyword search so identifiers do not need semantic retrieval.
  • Rerank selectively, not on every query.
  • Generate answers only when the user asks for synthesis.
  • Measure cost per search session, not per query.

How it works, step by step

  1. Inventory the corpus: document types, change frequency, permissions and languages.
  2. Hash documents and re-embed only changed or new content.
  3. Run hybrid retrieval with keyword search and vectors so exact terms stay cheap.
  4. Define query classes and apply reranking only to classes where it changes outcomes.
  5. Route answer synthesis to stronger tiers and leave result lists model-free.
  6. Filter by permissions before ranking and cache frequent query results where staleness is acceptable.
  7. Track cost per search session, click-through and answer usage by query class.
1Inventory thecorpus: documenttypes, change2Hash documents andre-embed onlychanged or new3Run hybridretrieval withkeyword search and4Define queryclasses and applyreranking only to5Route answersynthesis tostronger tiers and6Filter bypermissions beforeranking and cache

Try it yourself

Open the embedding cost calculator →

Embeddings scale with the corpus, not queries

In most enterprise search deployments, indexing is the largest controllable cost: every document chunk needs an embedding, and corpora are large. The mistake is re-embedding everything on a schedule. Hash-based incremental indexing means only new and changed content is embedded, and a corpus that changes one percent a day costs one percent a day to maintain.

Chunking strategy also affects index size. Smaller chunks mean more vectors and finer precision; larger chunks mean fewer vectors and better context per result. Tune chunk size against retrieval quality, then stop — index re-churn is expensive and rarely improves results.

Hybrid retrieval and selective reranking

Pure vector search is expensive when the query is an exact identifier: part numbers, error codes, clause references. Keyword search handles those at effectively zero model cost, so blend both and let the query type decide the weighting. This improves quality and removes a class of unnecessary embedding work.

  • Rerank only query classes with measurable quality gains; reranking everything doubles retrieval cost.
  • Cache popular query results with short expiry where content is stable.
  • Filter by permission before ranking so inaccessible documents never consume scoring or generation budget.
  • Keep query embeddings cached for repeated searches.

Answer generation and measurement

Generated answers are the premium feature, and they should be opt-in or triggered by query class, not applied to every search. Result lists need retrieval, not generation. Answer synthesis belongs on strong models for complex questions, delivered with citations so users can verify.

Measure cost per search session and per answer, alongside click-through, answer citation accuracy and abandoned searches. A cheap result list that users ignore is more expensive than a good one. Flat self-serve plans keep indexing and querying predictable — see the live pricing page for plans — and the free plan with plugsky-micro and plugsky-lite is enough to build the index and query pipeline before scaling.

Honest comparison

Search cost leverRouted enterprise search on PlugskyFrontier generation on every queryVector-only search
Index upkeepIncremental by hashFull re-embed schedulesFull re-embed schedules
Exact identifiersKeyword search at no model costSemantic search for everythingOften misses exact terms
RerankingSelective by query classOn every result setDisabled
Answer generationOn demand with citationsAlways onNot available
EconomicsCost per session and answerCost per queryCost per index refresh

Frequently asked questions

What dominates enterprise search cost?

Indexing, because every chunk needs an embedding and corpora are large. Incremental re-embedding of changed documents is usually the single largest saving available.

Should every query use semantic search?

No. Exact identifiers such as part numbers and clause references are better served by keyword search, which costs nothing in model calls. Hybrid retrieval gets both.

Is reranking always worth it?

Only where it measurably changes outcomes. Apply reranking to query classes that benefit and skip it elsewhere, because reranking every result set doubles retrieval cost.

When should search generate answers?

When the user asks a synthesis question or the query class benefits from a written answer. Result lists should not trigger generation on every search.

How do I keep the index fresh affordably?

Hash documents, re-embed only new and changed content, and remove retired documents. The batch endpoint for large refresh jobs is coming soon.

How do permissions affect cost?

Filter by permission before ranking so inaccessible content never consumes retrieval, reranking or generation budget, and log access decisions for audit.

How do I measure search value?

Cost per search session alongside click-through, citation accuracy on generated answers and abandoned-search rate. Cheap results nobody uses are not savings.

Can I start on a free plan?

Yes. plugsky-micro and plugsky-lite are on the free plan with no card, and the 14-day full-access trial covers stronger models for answer synthesis.