mirror of
https://github.com/hanzoai/ai.git
synced 2026-11-20 14:13:23 +00:00
getContextLength dispatched on substrings of the model name ("deepseek",
"v3", "32b", ...) and every model shipped after a branch was written fell
through it. deepseek-v4-pro matched nothing and got the 16384 fallback, so
api.hanzo.ai answered every prompt over 16K with
the token count: [30111] exceeds the model:
[deepseek-v4-pro]'s maximum token count: [16384]
against a model DO actually serves with an 87K window. glm-5.2 hit the same
16384. The cap was ours, not the upstream's: we refused the request before
DO ever saw it. A table that must be edited per model release is always one
release out of date, so it is gone (321 lines -> 52).
GetContextLength is now: whatever config declares, else a 128K floor. Nine
provider files called the private getContextLength directly and so bypassed
models.yaml entirely -- they now go through the one public entry point.
models.yaml carries the windows, taken from the live DO catalog rather than
guessed: 34 upstream leaves declared so every zen alias inherits its real
served window through the alias chain. Corrected glm-5.2, which claimed
1048576 ("GLM-5.2's real served window") when DO serves it at 262144.
Real windows now: zen5/best 256K (glm-5.2), zen5-pro 87K (deepseek-v4-pro),
zen5-flash 64K, zen5-ultra 1M (claude-opus-4.8 -- the only 1M tier we expose).
150 lines
4.7 KiB
Go
150 lines
4.7 KiB
Go
// Copyright 2023-2025 Hanzo AI Inc. All Rights Reserved.
|
|
// Portions Copyright 2023 The OpenAgent Authors. All Rights Reserved.
|
|
//
|
|
// Licensed under the Apache License, Version 2.0 (the "License");
|
|
// you may not use this file except in compliance with the License.
|
|
// You may obtain a copy of the License at
|
|
//
|
|
// http://www.apache.org/licenses/LICENSE-2.0
|
|
//
|
|
// Unless required by applicable law or agreed to in writing, software
|
|
// distributed under the License is distributed on an "AS IS" BASIS,
|
|
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
|
// See the License for the specific language governing permissions and
|
|
// limitations under the License.
|
|
|
|
package model
|
|
|
|
import (
|
|
"context"
|
|
"fmt"
|
|
"io"
|
|
"strings"
|
|
|
|
cohere "github.com/cohere-ai/cohere-go/v2"
|
|
cohereclient "github.com/cohere-ai/cohere-go/v2/client"
|
|
"github.com/hanzoai/ai/i18n"
|
|
)
|
|
|
|
// https://docs.cohere.com/docs/command-beta#whats-the-context-window-on-the-command-models
|
|
var CohereDefaultTemperature float64 = 0.75
|
|
|
|
type CohereModelProvider struct {
|
|
secretKey string
|
|
subType string
|
|
temperature float64
|
|
maxTokens int
|
|
verbose bool
|
|
stop []string
|
|
}
|
|
|
|
type ChatMessage struct {
|
|
Role string `json:"role"`
|
|
Message string `json:"message"`
|
|
User *string `json:"user,omitempty"`
|
|
}
|
|
|
|
func NewCohereModelProvider(subType string, secretKey string) (*CohereModelProvider, error) {
|
|
return &CohereModelProvider{
|
|
secretKey: secretKey,
|
|
subType: subType,
|
|
}, nil
|
|
}
|
|
|
|
// GetPricing returns the pricing of the model
|
|
// https://cohere.com/pricing
|
|
func (c *CohereModelProvider) GetPricing() string {
|
|
return `URL:
|
|
https://cohere.com/pricing
|
|
|
|
Generate Model:
|
|
|
|
| Model | Input Price (Per 1,000,000 tokens) | Output Price (Per 1,000,000 tokens) |
|
|
|---------------|------------------------------------|-------------------------------------|
|
|
| Command Light | $0.30 | $0.60 |
|
|
| Command | $1.00 | $2.00 |
|
|
|
|
Embed Model:
|
|
|
|
| Model | Cost (Per 1,000,000 tokens) |
|
|
|------------|-----------------------------|
|
|
| Default | $0.10 |
|
|
`
|
|
}
|
|
|
|
func (p *CohereModelProvider) calculatePrice(modelResult *ModelResult, lang string) error {
|
|
var inputPricePerThousandTokens, outputPricePerThousandTokens float64
|
|
switch p.subType {
|
|
case "command-light", "command-light-nightly":
|
|
inputPricePerThousandTokens = 0.0003
|
|
outputPricePerThousandTokens = 0.0006
|
|
case "command", "command-nightly":
|
|
inputPricePerThousandTokens = 0.001
|
|
outputPricePerThousandTokens = 0.002
|
|
default:
|
|
return fmt.Errorf("%s", fmt.Sprintf(i18n.Translate(lang, "embedding:calculatePrice() error: unknown model type: %s"), p.subType))
|
|
}
|
|
|
|
inputPrice := getPrice(modelResult.PromptTokenCount, inputPricePerThousandTokens)
|
|
outputPrice := getPrice(modelResult.ResponseTokenCount, outputPricePerThousandTokens)
|
|
modelResult.TotalPrice = AddPrices(inputPrice, outputPrice)
|
|
modelResult.Currency = "USD"
|
|
return nil
|
|
}
|
|
|
|
func (p *CohereModelProvider) QueryText(message string, writer io.Writer, chat_history []*RawMessage, prompt string, knowledgeMessages []*RawMessage, agentInfo *AgentInfo, lang string) (*ModelResult, error) {
|
|
client := cohereclient.NewClient(
|
|
cohereclient.WithToken(p.secretKey),
|
|
)
|
|
ctx := context.Background()
|
|
|
|
// if p.maxTokens > 0, use p.maxTokens, otherwise use model's default Maxtokens
|
|
maxTokens := GetContextLength(p.subType)
|
|
if strings.HasPrefix(message, "$CloudDryRun$") {
|
|
modelResult, err := getDefaultModelResult(p.subType, message, "")
|
|
if err != nil {
|
|
return nil, fmt.Errorf("%s", i18n.Translate(lang, "model:cannot calculate tokens"))
|
|
}
|
|
if maxTokens > modelResult.TotalTokenCount {
|
|
return modelResult, nil
|
|
} else {
|
|
return nil, fmt.Errorf("%s", i18n.Translate(lang, "model:exceed max tokens"))
|
|
}
|
|
}
|
|
generation, err := client.Generate(
|
|
ctx,
|
|
&cohere.GenerateRequest{
|
|
Prompt: prompt,
|
|
Temperature: &CohereDefaultTemperature,
|
|
MaxTokens: &maxTokens,
|
|
Model: &p.subType,
|
|
},
|
|
)
|
|
if err != nil {
|
|
return nil, err
|
|
}
|
|
if len(generation.Generations) == 0 {
|
|
return nil, fmt.Errorf("%s", i18n.Translate(lang, "model:no generations returned"))
|
|
}
|
|
|
|
output := generation.Generations[0].Text
|
|
resp := strings.Split(output, "\n")[0]
|
|
|
|
_, err = fmt.Fprint(writer, resp)
|
|
if err != nil {
|
|
return nil, err
|
|
}
|
|
|
|
promptTokenCount := int(*generation.Meta.BilledUnits.InputTokens)
|
|
responseTokenCount := int(*generation.Meta.BilledUnits.OutputTokens)
|
|
modelResult := &ModelResult{PromptTokenCount: promptTokenCount, ResponseTokenCount: responseTokenCount}
|
|
modelResult.TotalTokenCount = modelResult.ResponseTokenCount + modelResult.PromptTokenCount
|
|
|
|
err = p.calculatePrice(modelResult, lang)
|
|
if err != nil {
|
|
return nil, err
|
|
}
|
|
|
|
return modelResult, nil
|
|
}
|