Questa pagina è stata tradotta automaticamente. Leggi l’originale in inglese.

Feature

core/semantic-caching

Lets an AI gateway answer a prompt from the replies it stored earlier when the prompt is close enough in meaning to one it answered before, rather than calling the model again. The gateway turns each prompt into a vector with embedding_model, serves a stored reply whose prompt is at least similarity_threshold alike, and keeps a reply for ttl_seconds. scope decides who shares a cache. With key, every caller presenting the same key to the gateway is served from one cache, so an end user behind that key can be served a reply generated for another, tool calls included. The default, end_user, keeps each end user the gateway identifies apart. A match on meaning suits single prompts: on multi-turn or agentic traffic consecutive turns read alike, so a reply to an earlier turn can be served for a later one. It does not meet core/data-masking on one component. The developer attaches it in the Aggregate, with its inputs in values; it takes no infra. core/gateway/litellm, azure-api-management, apisix and traefik-hub realize it, the last three keeping scope key alone, apisix and traefik-hub for the callers each Release declares; litellm, apisix and traefik-hub keep the replies in a core/datastore/redis wired to the gateway, litellm requiring its administrative key. On core/gateway/kong-enterprise and higress every attachment halts, naming the caller separation each cache lacks, and on agentgateway, naming the reply cache it lacks; on any component but these seven an attachment halts naming the type, rendering nothing.

Provenienza

UfficialeFirmato da un’identità di firma che questo catalogo riconosce come editore ufficiale.
Digest del contenuto
sha256:bac81b4175f269b56678e2188d363229d5d8028c18c7928702a321bc90c751c1
Pubblicato
2026-11-10T10:17:27Z
Identità di firma
https://token.actions.githubusercontent.com/https://github.com/aggrega-ai/aggrega-hub/.github/workflows/catalog-publish-sign.yml@refs/heads/main

Firma

  • Ufficialeapplication/vnd.dev.sigstore.bundle.v0.3+json

    https://token.actions.githubusercontent.com/https://github.com/aggrega-ai/aggrega-hub/.github/workflows/catalog-publish-sign.yml@refs/heads/main

    sha256:f7c90cc58a5dd4682daeea38a0b260125ad4758a8f36a5b1811b67ff7cacdfb6

Riferimento a questa versione

oci://hub.aggrega.ai/feature/core/semantic-caching:1.0.0
Si applica a
gateway.aggrega.ai/intelligence=semantic
In conflitto con
—

Parametri dello sviluppatore

Valori che lo sviluppatore imposta quando aggancia questa funzionalità.

Estende BaseComponents#/schemas/BaseFeatureParameter

NomeTipoPredefinitoObbligatorioDescrizione
similarity_thresholdnumber (min 0, max 1)—sìHow alike a prompt must be to one answered before for the stored reply to be served: the cosine similarity of their two vectors, from 0 to 1, where 1 is identical. A higher value serves fewer and closer matches; which value suits depends on embedding_model.
embedding_modelstring (min 1)—sìThe model that turns a prompt into the vector compared, named as the gateway knows it: an alias of the model map on an AI gateway, or on Azure API Management the id of a backend the instance holds for an embeddings model. Each realization refuses a name its gateway does not publish wherever the build can see what the gateway publishes; an instance's backends are not in the build, so the Azure API Management realization takes the id as written.
ttl_secondsinteger (min 1)—sìHow long a stored reply stays usable, in seconds.
scopestring (key | end_user)end_usernoWho shares one cache. end_user, the default: each end user the gateway identifies has a cache of its own, so none is served a reply generated for another. key: every caller presenting the same key to the gateway shares one cache, so a reply generated for one end user behind that key can be served to another sending a similar prompt. Each realization says how its gateway identifies an end user and what it does with a request naming none, and refuses end_user where its gateway cannot keep end users apart.

Contratto di output

Valori che questa funzionalità pubblica verso i consumatori.

Nessun valore di output dichiarato.

Annotazioni

Nessuna annotazione dichiarata.

Relazioni

Archi tipizzati da e verso altre entità pubblicate. Un riferimento senza una destinazione pubblicata viene mostrato come arco interrotto anziché come link.

Riferimento al concetto Feature →

Il catalogo Aggrega — un indice aperto di componenti, funzionalità, ricette e dei plugin che li renderizzano.

Realizzato con il design system condiviso di Aggrega.

Privacy