Use Cases + Implementation

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

Semantic search cost splits between building the index and answering queries. Keep embeddings incremental, blend keyword search so exact terms never need vectors, and rerank only the query classes that benefit. Generation is optional: return ranked results by default and reserve answer synthesis on strong or long-context models for questions that genuinely need it.

Key facts

EmbeddingsLive embeddings API; re-embed only changed content by hash
RetrievalHybrid keyword plus vector search in your own store
Router modelplugsky-fusion escalates for answer generation (live)
RerankingApply selectively by query class
Models30+ models; cheap tiers for query understanding, strong for synthesis
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
RoadmapBatch re-embedding is coming soon; run bounded refresh jobs today

TL;DR

  • Index once per change, not once per schedule.
  • Blend keyword search so identifiers stay free.
  • Rerank selectively; measure the gain per query class.
  • Results need retrieval, not generation.
  • Cache frequent queries and repeated embeddings.

How it works, step by step

  1. Hash indexed content and re-embed only new or changed chunks.
  2. Store vectors with metadata for filtering and permission checks.
  3. Run hybrid retrieval: keyword first for exact terms, vectors for meaning.
  4. Define query classes and rerank only where quality measurably improves.
  5. Cache query embeddings for repeated searches.
  6. Route answer generation to strong or long-context models only when synthesis is requested.
  7. Track cost per search session alongside click-through and answer usage.
1Hash indexedcontent andre-embed only new2Store vectors withmetadata forfiltering and3Run hybridretrieval: keywordfirst for exact4Define queryclasses and rerankonly where quality5Cache queryembeddings forrepeated searches.6Route answergeneration tostrong or

Try it yourself

Open the embedding cost calculator →

Index cost is maintenance, not hardware

The recurring cost of semantic search is keeping the index current. Every changed chunk needs a new embedding, and teams that rebuild on a schedule pay for content that did not change. Hashing content and re-embedding only what moved turns a fixed periodic cost into one proportional to actual churn.

Chunking decisions also linger: smaller chunks multiply vector count and index size, larger chunks give coarser precision. Choose a chunk size that passes your retrieval evaluation, then leave it alone. Index churn from constant retuning is a cost with no user benefit.

Hybrid retrieval and selective reranking

Semantic search is not always the right tool. Exact identifiers, error codes, names and quoted phrases are better served by keyword search, which costs nothing in model calls. A hybrid approach routes each query to the mechanism that suits it, and improves results while reducing embedding work.

  • Keep query embeddings cached; repeated searches are common in internal tools.
  • Filter by permission and metadata before ranking so irrelevant content never consumes budget.
  • Rerank only query classes with measured gains; blanket reranking doubles retrieval cost.
  • Log which mechanism answered each query to tune the balance over time.

Answers on request, not by default

Generated answers are a feature, not a default. Result lists should cost retrieval only; synthesis runs when the user asks a question that needs it. When enabled, route factual questions to cheap tiers and multi-source synthesis to stronger or long-context models, always with citations.

Measure cost per search session, answer usage and citation accuracy. Start on the free plan with plugsky-micro and plugsky-lite to build the retrieval and generation split, then evaluate stronger models for synthesis during the 14-day full-access trial; plans are on the live pricing page.

Honest comparison

Search cost leverRouted semantic searchVectors for every queryFull re-embed on schedule
Index upkeepIncremental by hashSameFixed periodic cost
Exact identifiersKeyword search, no model callEmbedding call per queryDepends on index
RerankingSelective by query classAlways onDepends on index
AnswersOn request with citationsOften always onDepends on index
EvidenceCost per session and answerCost per queryCost per refresh

Frequently asked questions

What is the largest semantic search cost?

Index maintenance on changing corpora and answer generation on high query volume, depending on your design. Incremental embedding and on-request answers are the two biggest levers.

Do I still need keyword search?

Yes. Exact identifiers, names and quoted phrases are handled better and more cheaply by keyword search. Hybrid retrieval improves quality and cuts embedding work.

Should every query be reranked?

No. Rerank the query classes where you can measure an improvement. Running a reranker on all queries doubles retrieval work for inconsistent benefit.

When should search generate an answer?

When the user asks a synthesis question. Result lists should return ranked documents without a generation call, so cost stays proportional to intent.

How do I keep the index fresh?

Hash content and re-embed only changed chunks, and remove retired documents. Bounded refresh jobs run today; batch endpoints are coming soon.

Can I cache search results?

Cache query embeddings always, and cache results or answers for frequent queries with a time-to-live tied to content change.

How do I evaluate the split?

Track click-through, citation accuracy on generated answers and cost per session by query class, and compare mechanisms to tune the hybrid weighting.

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.