Cloudflare models#

Cloudflare Workers AI is an ordinary Antfly provider. Configure account_id, the full @cf/owner/model or @hf/owner/model name, and an api_key secret reference. Credentials default to CLOUDFLARE_API_TOKEN. Optional gateway_id uses cf-aig-gateway-id; credentials, pools, quotas and deadlines stay in the existing provider runtimes. Native library callers supply resolved tokens.

Cloudflare does not expose a standalone chunking model matching Antfly's chunk contract. Fixed-label sentiment and image classifiers do not implement arbitrary caller labels, and VAD/end-of-turn detection does not imply timed audio segments. Sparse embeddings, image embedding models without a qualified request schema, Cloudflare Markdown conversion and tabular predictors are not advertised by these adapters. Document agents retain Antfly's PDF/PPTX source handling and render pages only when their policy needs visual reading. Generated vision output is marked interpretation; the adapter does not invent OCR boxes or source citations.

Configuration#

{
  "embedders": {
    "cf-text": {
      "provider": "cloudflare",
      "account_id": "account",
      "model": "@cf/baai/bge-base-en-v1.5",
      "dimensions": 768,
      "api_key": "${secret:cloudflare.api_token}"
    }
  },
  "generators": {
    "cf-chat": {
      "provider": "cloudflare",
      "account_id": "account",
      "model": "@cf/meta/llama-3.3-70b-instruct-fp8-fast",
      "api_key": "${secret:cloudflare.api_token}",
      "max_tokens": 1024
    }
  },
  "rewriters": {
    "clean-text": {
      "generator": "cf-chat",
      "instructions": "Rewrite clearly while preserving meaning. Return only the rewritten text."
    }
  },
  "deciders": {
    "route-by-similarity": {
      "decision_method": "embedding_similarity",
      "embedder": "cf-text",
      "embedding_options": { "dimensions": 768, "min_similarity": 0.5 }
    },
    "generated-choice": {
      "decision_method": "generated",
      "generator": {
        "provider": "cloudflare",
        "account_id": "account",
        "model": "@cf/meta/llama-3.3-70b-instruct-fp8-fast",
        "api_key": "${secret:cloudflare.api_token}"
      },
      "probabilities": "verdict_only"
    }
  }
}

Embedding response dimensions are validated, and fixed Cloudflare models are not asked to resize vectors. Text requests are batched within the native 100-item ceiling and retain one quota observation per HTTP attempt. Similarity decisions batch and deduplicate input/category embeddings, normalize example centroids, and return cosine scores with explicit tie/threshold abstentions. Their token accounting uses a conservative byte upper bound marked estimated: true, because the ordinary embedding interface does not expose measured token usage. Hosted routes do not reuse fitted local model calibrations or claim immutable weights.

Generated extraction and decisions validate output through the existing typed contracts. Qualified Cloudflare models use tool calls with explicit probabilities: "verdict_only"; they do not claim unavailable token logprobs. Model support for tools, images and structured output remains model-specific. Clairic's native decimal precision is limited to 0–6. Whisper rejects diarization; MP3 synthesis rejects unsupported speed and container overrides before dispatch.

Clairic validates the effective extraction schema version (1 or 2) before dispatch. One extraction deadline covers input validation and all rows in a serial batch; each upstream request receives only the remaining time. Explicit request controls take precedence over the shared client's timeout and cancellation defaults. Malformed successful vision responses are upstream failures, reported as HTTP 502 by document endpoints.

See declarative document models for whole PDF, slide and receipt sources and bounded reading/extraction policies. The provider wire types are generated from pinned official model schemas; transport fixtures exercise the adapters without consuming live model credits. Live model quality and account-specific availability require separate qualification.

Ordinary hosted vision reads require an explicit successful finish reason. Cloudflare/OpenAI-compatible length responses fail with ReadOutputTruncated; filtered, refused, tool-only, missing or unknown endings remain unresolved. Vertex uses the same acceptance policy for STOP and MAX_TOKENS, concatenates visible text parts in order, and excludes thought text. Hitting the requested token count alone does not imply truncation.

A single response adapter preserves measured prompt/completion usage on each reader item and on the reported batch. Missing usage stays absent; explicit zero counts stay measured zero. Malformed, overflowing or inconsistent counters are rejected rather than coerced or estimated. For Vertex, tool-result prompt tokens count as input, and thought tokens count as output, matching the provider's usage categories. Text is copied into owned result storage without intermediate JSON buffers, and incomplete ordinary results release all response allocations before returning.

Materialized similarity decisions#

Persist a named embedder's configuration with the decision producer. The embedders map freezes the dependency used for replay; it does not resolve against a node's current named-model configuration. Changing that dependency changes the specification hash and follows the normal enrichment rebuild lifecycle. Credentials remain secret references or provider environment defaults.

{
  "version": "v1",
  "decider": {
    "decision_method": "embedding_similarity",
    "embedder": "cf-text",
    "embedding_options": { "dimensions": 768, "min_similarity": 0.5 }
  },
  "embedders": {
    "cf-text": {
      "provider": "cloudflare",
      "account_id": "account",
      "model": "@cf/baai/bge-base-en-v1.5",
      "dimensions": 768,
      "api_key": "${secret:cloudflare.api_token}"
    }
  },
  "questions": [{
    "name": "route",
    "type": "choice",
    "instructions": "Choose the category that best describes the input.",
    "choices": [{ "value": "receipt" }, { "value": "slide" }]
  }]
}

Compatible rows share bounded embedding work and category prototypes. An invalid query vector fails its own row; valid siblings remain available, and the batch's conservative token estimate is retained even if scoring or dispatch fails. SQL and HTTP execution borrow service-owned provider clients. Standalone enrichment owners retain their own provider runtime across batches. The batch's embedding_options.task_type selects CLUSTERING or CLASSIFICATION. Task-aware providers receive that task through both linked and hosted execution; retrieval-specific overrides do not replace it. The task participates in prototype provenance while all compatible rows and questions share one deduplicated embedding call. Legacy linked callbacks without task context reject these decisions before inference.

Generator-backed rewriting requires a complete generation result. Embedded Antfly generators use the JSON callback to preserve finish reasons and usage; legacy text-only callbacks are rejected before execution for this task.