Files
ai/model/huggingface.go
zeekay ea1f8248eb model: rip the context-length guessing table; config is the only truth
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release out of date, so it is gone (321 lines -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).
2026-07-12 17:19:26 -07:00

103 lines
2.9 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"
"github.com/hanzoai/ai/i18n"
"github.com/hanzoai/ai/proxy"
"github.com/hupe1980/go-huggingface"
)
type HuggingFaceModelProvider struct {
subType string
secretKey string
temperature float32
}
func NewHuggingFaceModelProvider(subType string, secretKey string, temperature float32) (*HuggingFaceModelProvider, error) {
return &HuggingFaceModelProvider{subType: subType, secretKey: secretKey, temperature: temperature}, nil
}
func (p *HuggingFaceModelProvider) GetPricing() string {
return `URL:
https://huggingface.co/pricing
Not charged
`
}
func (p *HuggingFaceModelProvider) calculatePrice(modelResult *ModelResult) error {
modelResult.Currency = "USD"
return nil
}
func (p *HuggingFaceModelProvider) QueryText(question string, writer io.Writer, history []*RawMessage, prompt string, knowledgeMessages []*RawMessage, agentInfo *AgentInfo, lang string) (*ModelResult, error) {
ctx := context.Background()
client := huggingface.NewInferenceClient(p.secretKey, func(o *huggingface.InferenceClientOptions) {
o.HTTPClient = proxy.ProxyHttpClient
})
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"))
}
}
resp, err := client.TextGeneration(ctx, &huggingface.TextGenerationRequest{
Inputs: question,
Parameters: huggingface.TextGenerationParameters{
Temperature: huggingface.PTR(float64(p.temperature)),
},
Options: huggingface.Options{
WaitForModel: huggingface.PTR(true),
},
Model: p.subType,
})
if err != nil {
return nil, err
}
respText := strings.Split(resp[0].GeneratedText, "\n")[0]
_, err = fmt.Fprint(writer, respText)
if err != nil {
return nil, err
}
modelResult, err := getDefaultModelResult(p.subType, question, respText)
if err != nil {
return nil, err
}
err = p.calculatePrice(modelResult)
if err != nil {
return nil, err
}
return modelResult, nil
}