Files
HomeAgent/pkg/embedding/embedding.go
JianFeeeee f256098c9f refactor: 全面改名为 HologramAgent + 提示词技术内容下沉到 skill/工具描述 + 修 21 处乱码
## ① 全面改名 HomeAgent → HologramAgent

仓库已改名,但代码与文档里仍有大量旧名。本次按"用户能看到的优先"全量替换:

- 用户可见:登录页标题/页脚、默认人格第一句、CLI 状态输出、启动日志、
  plugin_install 的工具描述、AI 图片插件的 User-Agent
- 文档与脚本:README、assets/docs、docs/zh、deploy/*.sh、.github/workflows
- 鸿蒙工程:cmd/ohos/HomeAgent → cmd/ohos/HologramAgent(git mv)
  目录内的 module/app 描述同步;**证书路径原样保留**
  (default_HomeAgent_xxx.cer/p7b/p12 是签名材料的实际文件名,改了会找不到证书)
- 插件元数据:plugin.json 的 author

★ 必须保持原样的两处(改了会坏):
  · legacySeededPromptSignature = "你是 HomeAgent 的看板娘「小宅」…"
    —— 那是识别历史播种人格卡的**原文锚点**,改了就识别不出旧数据
  · 公开 SDK 的 ProxyAuthHomeAgent = "homeagent"
    —— 那是 SDK 契约里的值,属独立仓库,本仓不改

## ② 修 21 处乱码(U+FFFD)

这些是历史编辑留下的编码破坏,全在注释里,例如:
  「赌注重下���遍」→「赌注重下一遍」
  「「不��滤」」  →「「不过滤」」
  「**待��退场**」→「**待退场**」

★ 保留 providers/chineseclip/tokenizer.go 里的 2 处:那是**故意**用
  U+FFFD 演示"非法 UTF-8 字节被归一成 U+FFFD",不是乱码。

## ③ 提示词技术内容下沉(用户:技术性内容应该收敛到 skill 与 -help 中)

实测系统提示词 2309 字,其中两段几乎全是**模型无法据以行动的实现细节**:

  【工具执行顺序】1272 字(ParallelSafe 声明、批内降级、batchRunnable 判据…)
  【三层记忆】   1069 字(L1/L2/L3 编号、消费式删除、向量空间、指纹…)

收敛后 **2309 → 1723 字(-25%)**,保留的全是"模型据此做什么不同的事":

  工具顺序:默认并行 / 不要依赖先后 / 同通道保序 / 写类工具可能串行
  三层记忆:时序里能看最近 / doc_query 把历史事件按原始时间戳插回时序 /
           关系用 memory_recall / 成文知识用 knowledge_search

判断口径写进判据:新加一段提示词前先问"模型会据此做什么不同的事",
答不上来就该进 skill 或工具描述。

## 判据

prompt_memory_test.go 新增 2 条:
- TestSystemPromptHasNoImplementationJargon
  黑名单:ParallelSafe / batchRunnable / 消费式 / DenseCosine / fingerprint /
  向量空间 / 批内降级 / 自激
- TestSystemPromptStaysCompact:上限 2000 字(当前 1723)

变异:把 "ParallelSafe 声明、批内降级" 加回工具顺序段 ⇒ 第一条立刻红。

## 验证

- go build ./... / go vet ./... 干净
- go test ./... -count=1:58 包 0 失败
- make check-client-versions:客户端与内核版本一致
2026-10-22 20:49:44 +08:00

173 lines
5.1 KiB
Go

// Package embedding defines the public SPI for dense multimodal embedding providers.
//
// The HologramAgent core depends only on this package. Model runtimes, tokenizers,
// preprocessing, media decoding, and model-specific configuration belong in
// provider packages registered with Register.
package embedding
import (
"context"
"errors"
"fmt"
"math"
"sort"
"strings"
"sync"
)
// Modality identifies the semantic kind of an embedding input. Providers may
// support additional modality strings; the constants are the common values.
type Modality string
const (
ModalityText Modality = "text"
ModalityImage Modality = "image"
ModalityAudio Modality = "audio"
ModalityVideo Modality = "video"
)
// Purpose tells a provider how the vector will be used. Providers whose model
// distinguishes query and document prompts can map this value accordingly.
type Purpose string
const (
PurposeQuery Purpose = "query"
PurposeDocument Purpose = "document"
)
// Input is the model-neutral request passed to a provider.
//
// Data is deliberately opaque to the core. MIME describes the encoding; the
// selected provider owns decoding, frame sampling, preprocessing, and all
// other model-specific interpretation. Text is used for textual inputs.
type Input struct {
Modality Modality
Purpose Purpose
Text string
Data []byte
MIME string
Metadata map[string]string
}
// Info describes one vector space. Fingerprint must change whenever vectors
// cease to be comparable with vectors produced by a previous provider build.
type Info struct {
Dimension int
Fingerprint string
Modalities []Modality
}
// Provider is the public Go extension point for a dense multimodal vector
// space. Implementations must be safe for concurrent Embed calls unless their
// factory documents otherwise and serializes internally.
type Provider interface {
Embed(context.Context, Input) ([]float64, error)
Info() Info
Close()
}
// Config contains provider-owned options. The core does not interpret option
// names or values; it only passes core.memory.multimodal_space.options.*
// through after stripping the prefix.
type Config struct {
Options map[string]string
}
// Factory constructs a provider instance.
type Factory func(Config) (Provider, error)
var (
// ErrUnsupportedModality means this vector space has no native encoder for
// the requested modality. Callers must not substitute another model's vector.
ErrUnsupportedModality = errors.New("embedding: unsupported modality")
registryMu sync.RWMutex
registry = make(map[string]Factory)
)
// Register makes a provider factory available under name. It is normally
// called from a provider package's init function. Duplicate names panic so a
// build cannot silently select whichever package initialized last.
func Register(name string, factory Factory) {
name = strings.TrimSpace(name)
if name == "" {
panic("embedding: register empty provider name")
}
if factory == nil {
panic("embedding: register nil factory for " + name)
}
registryMu.Lock()
defer registryMu.Unlock()
if _, exists := registry[name]; exists {
panic("embedding: provider already registered: " + name)
}
registry[name] = factory
}
// Open constructs a registered provider and validates its vector-space identity.
func Open(name string, cfg Config) (Provider, error) {
name = strings.TrimSpace(name)
registryMu.RLock()
factory := registry[name]
registryMu.RUnlock()
if factory == nil {
return nil, fmt.Errorf("embedding: unknown provider %q (available: %s)", name, strings.Join(Names(), ", "))
}
provider, err := factory(cloneConfig(cfg))
if err != nil {
return nil, fmt.Errorf("embedding: open provider %q: %w", name, err)
}
if provider == nil {
return nil, fmt.Errorf("embedding: provider %q returned nil", name)
}
if err := ValidateInfo(provider.Info()); err != nil {
provider.Close()
return nil, fmt.Errorf("embedding: provider %q: %w", name, err)
}
return provider, nil
}
// Names returns registered provider names in deterministic order.
func Names() []string {
registryMu.RLock()
defer registryMu.RUnlock()
names := make([]string, 0, len(registry))
for name := range registry {
names = append(names, name)
}
sort.Strings(names)
return names
}
// ValidateInfo checks the stable identity required by vector persistence.
func ValidateInfo(info Info) error {
if info.Dimension <= 0 {
return fmt.Errorf("invalid dimension %d", info.Dimension)
}
if strings.TrimSpace(info.Fingerprint) == "" {
return errors.New("empty fingerprint")
}
return nil
}
// ValidateVector rejects malformed provider output before it reaches storage.
func ValidateVector(vec []float64, dimension int) error {
if len(vec) != dimension {
return fmt.Errorf("embedding: vector dimension %d, want %d", len(vec), dimension)
}
for i, value := range vec {
if math.IsNaN(value) || math.IsInf(value, 0) {
return fmt.Errorf("embedding: vector value %d is not finite", i)
}
}
return nil
}
func cloneConfig(cfg Config) Config {
out := Config{Options: make(map[string]string, len(cfg.Options))}
for key, value := range cfg.Options {
out.Options[key] = value
}
return out
}