Subprocesadores

Actualizado al: 2026-10-22

Estos son los terceros que pueden recibir contenido derivado de tu uso de LexI. La lista se genera desde el registro que también gobierna el código, de modo que un proveedor nuevo no puede entrar al producto sin aparecer aquí. Ver el aviso de privacidad.

Mientras cualquier fila abajo diga “Sin verificar”, no afirmamos que ningún Proveedor de IA entrena con tus datos a nivel de todo el sistema. La página principal publica una versión más acotada de esa promesa, limitada a los proveedores cuya postura sí está verificada.

ProveedorTipoQué recibeEntrenamientoRetenciónRegiónBase
Cohere Inc.Proveedor de IAquery_text, corpus_chunksSin verificar — no afirmamos nada todavíaunknownUScohere.com/terms-of-use, read 2026-08-25 — public ToU grants improvement + third-party-sharing rights over API data (clause quoted in training_citation); whether a signed agreement displaces it is unrecorded (CROSS-ADR-0024, DQ10) “IMPROVE AND ENHANCE THE COHERE SOLUTION...BY SHARING API DATA AND FINETUNING DATA WITH THIRD PARTIES.” (fuente, consultado el 2026-08-25)
Google LLCProveedor de IAprompt, document_textNo entrena con tus datosunknownUSai.google.dev/gemini-api/terms, read 2026-08-25 — paid tier: Google does not use prompts or responses to improve its products. Tier RECORDED 2026-08-28: the owner confirmed this key's Google Cloud project is on the paid (billing-enabled) tier (CROSS-ADR-0024, DQ10 — Google half closed) “Unpaid Services: "Google uses the content you submit to the Services and any generated responses to provide, improve, and develop Google products and services" — "Human reviewers may read, annotate, and process your API input and output." Paid Services: "Google doesn't use your prompts (including associated system instructions, cached content, and files such as images, videos, or documents) or responses to improve our products".” (fuente, consultado el 2026-08-25)
Google LLCProveedor de IAdocument_textNo entrena con tus datosnone for the online :process call we make — processed in memory, encrypted in flight, not persisted to disk (vendor doc). The batch endpoint, which this repo does not call, stores briefly under an ephemeral key.UScloud.google.com/document-ai/docs/security, read 2026-09-02 — Google does not use Document AI content for any purpose except providing the service, and never to train its Document AI models (clause quoted in training_citation) “No. Google does not use any of your content (such as documents and predictions) for any purpose except to provide you with the Document AI service.” (fuente, consultado el 2026-09-02)
OpenAI, L.L.C.Proveedor de IAprompt, document_textNo entrena con tus datosabuse-monitoring logs up to 30 days (vendor default)USdevelopers.openai.com/api/docs/guides/your-data, read 2026-08-25 — API data is not used for training absent an explicit opt-in this repo does not enable (clause quoted in training_citation) “data sent to the OpenAI API is not used to train or improve OpenAI models (unless you explicitly opt in to share data with us)” (fuente, consultado el 2026-08-25)
Anthropic, PBCProveedor de IAprompt, document_textNo entrena con tus datosunknownUSanthropic.com/legal/commercial-terms Section B, read 2026-08-25 — "Anthropic may not train models on Customer Content from Services" (clause quoted in training_citation) “Anthropic may not train models on Customer Content from Services.” (fuente, consultado el 2026-08-25)
Vercel Inc.Proveedor de IAprompt, document_text, query_text, corpus_chunksNo entrena con tus datosModel legs (chat, chat-brief, contract-analysis, entity-extraction, router and retrieval-plan slots — every LangChain model built on the gateway): zero data retention APPLIED, with ONE exception below (D95) — the code sends `zeroDataRetention: true` (GATEWAY_CHAT_PROVIDER_OPTIONS, D89). One live call per slot candidate for chat, contract-analysis, entity-extraction and router (2026-10-21, docs/eval/zdr-rollout-2026-10-21/zdr-chat-slots-2026-10-21.json) returned `enabledZeroDataRetention: true`; with ZDR requested the gateway drops our stored BYOK attempt and plans only ZDR-capable providers. chat-brief and retrieval-plan use the same constant and were not captured separately. The exception, per plan (D95 as amended 2026-10-22; `answer_route`): on the FREE plan, every chat-graph model call of a turn on the chat, chat-brief and entity-extraction slots is sent WITHOUT `zeroDataRetention`, pinned `only: ["anthropic"]` (FREE_GATEWAY_PROVIDER_OPTIONS). The gateway serves it on our stored BYOK Anthropic key, under Anthropic's standard API retention — see the `anthropic` upstream (live proof 2026-10-22: docs/eval/zdr-rollout-2026-10-21/d92-free-byok-live-2026-10-22.json). When that key's credit is exhausted the gateway serves the call on its own system credential — still Anthropic, still WITHOUT ZDR, billed as gateway credit; the code reports each such turn (`answer_route_byok_fallthrough`). Paid plans and turns whose plan could not be read always get ZDR. Combined-router legs (router-jev, router-tiebreak): zero data retention APPLIED — the code sends `zeroDataRetention: true`; on one live call per leg (2026-10-21; see `upstreams`) the gateway accepted it and reported every planned provider on the leg as ZDR-capable. Only the first provider served each call (typesafe-ai, fireworks); groq and bedrock rest on that planning report. The Jev endpoint returns no `enabledZeroDataRetention` field; the tie-breaker returned `true`. Every other gateway leg (embed, rerank, document extraction (vision)): ZDR NOT requested. The account was on the Hobby plan, where `zeroDataRetention: true` returned HTTP 403 (measured 2026-08-29); it is on Pro as of 2026-10-21, and applying ZDR to those legs is a separate owner decision (D32 follow-up). Of those legs, only the retrieval slugs were checked in the catalogue (2026-08-29): the embed slug is zdr:"some", both v4 rerank slugs are zdr:"none" (only cohere/rerank-v3.5 is "all"), so the v4 rerank leg could not obtain ZDR; the document-extraction (vision) slug was not checked.USGET https://ai-gateway.vercel.sh/v1/models — no_training: "all", live catalogue read 2026-08-24; re-read 2026-08-29 covering the retrieval slugs specifically: cohere/embed-v4.0 → type=embedding no_training=all zdr=some regions=[us] pricing.input=0.00000012; cohere/rerank-v4-pro → type=reranking no_training=all zdr=none pricing={} (2026-08-29). Note the asymmetry: the embed slug is zdr:"some" while both v4 rerank slugs are zdr:"none" (only cohere/rerank-v3.5 is "all"). Observed 2026-08-29, while the account was on Hobby; on Pro (since 2026-10) it still means the v4 rerank leg cannot obtain ZDR — a fact the D32 follow-up needs. “AI Gateway does not use your prompts or responses for training purposes. Your data is processed solely to fulfill your requests and is not retained for model improvement.” (fuente, consultado el 2026-08-25)
Amazon Web Services, Inc.Proveedor de IAprompt, query_text, corpus_chunksNo entrena con tus datosunknownUSaws.amazon.com/bedrock/faqs, read 2026-08-25 — content is not used to improve base models and is not shared with model providers (clause quoted in training_citation) “With Amazon Bedrock, your content is not used to improve the base models and is not shared with any model providers.” (fuente, consultado el 2026-08-25)
Zilliz Inc.Infraestructuraembedding_vector, enum_filtersNo aplica (no recibe texto para modelos)cluster storage, indefiniteUSsigned DPA; receives no free text — verified in code 2026-08-25 (mexican-legal-search.ts:710)
Langfuse GmbHInfraestructuramasked_traceNo aplica (no recibe texto para modelos)Langfuse project setting — owner input openEUCROSS-ADR-0023 — fenced documents masked to a character count; Preview only, production gated on LANGFUSE_ENABLE_PRODUCTION
Google LLCInfraestructuraaccount, document_text, chat_messagesNo aplica (no recibe texto para modelos)see docs/living/DATA-RETENTION.md rows 1-11USus-central1, encrypted at rest and in transit — DATA-RETENTION.md, verified 2026-08-24
Vercel Inc.Infraestructurarequest_metadataNo aplica (no recibe texto para modelos)Vercel plan default (unverified)USDATA-RETENTION.md row 12 — logs carry counts and sizes, no filenames, no document text

Quién genera las respuestas, según tu plan

En el plan gratuito aplica a las llamadas del chat que usan el modelo de respuesta (clasificar y preparar la consulta, pedir aclaraciones y generar la respuesta) y a la extracción de datos para cálculos. El enrutador de modelos, cuando está activo, y los resúmenes de la conversación y de los adjuntos usan retención cero (ZDR) en todos los planes. El plan gratuito incluye las cuentas de prueba; un asiento de equipo cuenta como plan de pago. Lo demás se procesa como indica la tabla anterior. Verificado el 2026-10-22.

A través de Vercel

Proveedores a los que el AI Gateway de Vercel reenvía tu consulta: para clasificarla y, en el plan gratuito, para responderla.

ProveedorPara quéQué recibeEntrenamientoRetenciónBase
TypeSafe AI, Inc.Clasificación de la consultaTu consulta y el historial recienteNo entrena con tus datos, según el listado de VercelRetención cero (ZDR) aplicada a través de VercelListado de endpoints de Vercel AI Gateway, consultado el 2026-10-21; retención cero solicitada y aceptada por Vercel en una llamada en vivo el 2026-10-21.
Fireworks.ai, Inc.Desempate en consultas dudosasTu consulta y el historial recienteNo entrena con tus datos, según el listado de VercelRetención cero (ZDR) aplicada a través de VercelListado de endpoints de Vercel AI Gateway, consultado el 2026-10-21; retención cero solicitada y aceptada por Vercel en una llamada en vivo el 2026-10-21.
Groq LLCDesempate en consultas dudosas (respaldo)Tu consulta y el historial recienteNo entrena con tus datos, según el listado de VercelRetención cero (ZDR) aplicada a través de VercelListado de endpoints de Vercel AI Gateway, consultado el 2026-10-21; retención cero solicitada y aceptada por Vercel en una llamada en vivo el 2026-10-21.
Amazon Web Services, Inc. (Amazon Bedrock, vía la cuenta de Vercel, no la nuestra)Desempate en consultas dudosas (respaldo)Tu consulta y el historial recienteNo entrena con tus datos, según el listado de VercelRetención cero (ZDR) aplicada a través de VercelListado de endpoints de Vercel AI Gateway, consultado el 2026-10-21; retención cero solicitada y aceptada por Vercel en una llamada en vivo el 2026-10-21.
Anthropic, PBCRespuestas del plan gratuitoTu consulta, el historial, el texto de tus adjuntos y el contexto legal recuperadoNo entrena con tus datos, según el listado de VercelRetención estándar de la API de Anthropic (Anthropic indica que borra entradas y salidas en 30 días, con excepciones); sin retención cero (ZDR)Listado de endpoints de Vercel AI Gateway, consultado el 2026-10-22.