contextcount.Estimator uses the request’s API format to choose one of the bundled request encoders. Its zero value is ready for use.

Selection

switch req.Model.APIFormat {
case model.APIFormatOpenAI:
	body, err = openaiapi.EncodeRequest(req, false)
case model.APIFormatOpenAIResponses:
	body, err = openairesponses.EncodeRequest(req, false)
case model.APIFormatAnthropic:
	body, err = anthropicapi.EncodeRequest(req, false)
case model.APIFormatGemini:
	body, err = geminiapi.EncodeRequest(req)
case model.APIFormatBedrockConverse:
	body, err = bedrockconverse.EncodeRequest(req)
default:
	return nil, &contextcount.UnsupportedAPIFormatError{APIFormat: req.Model.APIFormat}
}

Bedrock Converse is in the bundle: Converse and ConverseStream share one request body, so the encoder takes no mode and the estimator counts exactly what inference sends. A format with no bundled encoder returns a typed UnsupportedAPIFormatError, not a guessed count.

Formula

The estimator counts encoded bytes with a ceiling division by four:

estimatedTokens = ceil(len(encodedRequest) / 4).

The revision constant bundled-openai-responses-anthropic-gemini-bedrock-request-bytes-div4-v3 identifies the encoder suite and formula. A count-affecting codec change requires a new revision so stored measurements remain attributable.

Source and proof

Run go test ./contextcount.