Files
zeekay d2f91f1c9a feat(model): configurable per-model context windows via models.yaml
context_window is now a first-class per-model field in models.yaml — the
single source of truth for max context, settable per provider/model without
a rebuild. This decomplects context length from the hardcoded getContextLength
Go table (which only knew legacy families and fell through to a 16K fallback
for every modern model: glm-5.x, kimi-k2.x, deepseek-v4, qwen3.5).

- ModelDef.ContextWindow + modelRoute.contextWindow (yaml -> route)
- ModelConfig.ContextWindow(model): config-first lookup that follows the
  alias chain — zen5 (no own window) inherits glm-5.2's declared 1M
- model.SetContextWindowResolver: a package-var hook controllers wires at
  init (no model->controllers import cycle); nil = table-only fallback
- GetContextLength: resolver-first, table-fallback entry point
- do-ai path (openai.go, chatglm.go) uses GetContextLength so glm-5.2/zen5
  get the real 1M window instead of the 16K fallback
- models.yaml: declare context_window: 1048576 on glm-5.2 (zen5 inherits it)
- tests: TestContextWindow locks in direct + alias-chain + unset/unknown cases
2026-07-12 12:31:43 -07:00

125 lines
3.7 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 (
"fmt"
"io"
"net/http"
"strings"
"time"
"github.com/hanzoai/ai/i18n"
"github.com/leverly/ChatGLM/client"
)
type ChatGLMModelProvider struct {
subType string
clientSecret string
}
func NewChatGLMModelProvider(subType string, clientSecret string) (*ChatGLMModelProvider, error) {
return &ChatGLMModelProvider{subType: subType, clientSecret: clientSecret}, nil
}
func (c *ChatGLMModelProvider) GetPricing() string {
return `URL:
https://open.bigmodel.cn/pricing
Generate Model:
| Model | Context Length | Unit Price (Per 1,000 tokens) |
|--------------|----------------|-------------------------------|
| GLM-3-Turbo | 128K | 0.005 yuan / K tokens |
| GLM-4 | 128K | 0.1 yuan / K tokens |
| GLM-4V | 2K | 0.1 yuan / K tokens |
`
}
func (p *ChatGLMModelProvider) calculatePrice(modelResult *ModelResult, lang string) error {
price := 0.0
switch p.subType {
case "glm-3-turbo":
price = getPrice(modelResult.TotalTokenCount, 0.005)
case "glm-4", "glm-4v":
price = getPrice(modelResult.TotalTokenCount, 0.1)
default:
return fmt.Errorf("%s", fmt.Sprintf(i18n.Translate(lang, "embedding:calculatePrice() error: unknown model type: %s"), p.subType))
}
modelResult.TotalPrice = price
modelResult.Currency = "CNY"
return nil
}
func (p *ChatGLMModelProvider) QueryText(question string, writer io.Writer, history []*RawMessage, prompt string, knowledgeMessages []*RawMessage, agentInfo *AgentInfo, lang string) (*ModelResult, error) {
proxy := client.NewChatGLMClient(p.clientSecret, 30*time.Second)
messages := []client.Message{{Role: "user", Content: question}}
taskId, err := proxy.AsyncInvoke(p.subType, 0.2, messages)
if err != nil {
return nil, err
}
flusher, ok := writer.(http.Flusher)
if !ok {
return nil, fmt.Errorf("%s", i18n.Translate(lang, "model:writer does not implement http.Flusher"))
}
flushData := func(data string) error {
if _, err = fmt.Fprintf(writer, "event: message\ndata: %s\n\n", data); err != nil {
return err
}
flusher.Flush()
return nil
}
if strings.HasPrefix(question, "$CloudDryRun$") {
modelResult, err := getDefaultModelResult(p.subType, question, "")
if err != nil {
return nil, fmt.Errorf("%s", i18n.Translate(lang, "model:cannot calculate tokens"))
}
if GetContextLength(p.subType) > modelResult.TotalTokenCount {
return modelResult, nil
} else {
return nil, fmt.Errorf("%s", i18n.Translate(lang, "model:exceed max tokens"))
}
}
response, err := proxy.AsyncInvokeTask(p.subType, taskId)
if err != nil {
return nil, err
}
content := (*response.Choices)[0].Content
err = flushData(content)
if err != nil {
return nil, err
}
modelResult := &ModelResult{}
modelResult.PromptTokenCount = response.Usage.PromptTokens
modelResult.ResponseTokenCount = response.Usage.CompletionTokens
modelResult.TotalTokenCount = response.Usage.TotalTokens
err = p.calculatePrice(modelResult, lang)
if err != nil {
return nil, err
}
return modelResult, nil
}