1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
|
package rag
import (
"bytes"
"encoding/json"
"errors"
"fmt"
"gf-lt/config"
"gf-lt/models"
"log/slog"
"net/http"
"github.com/takara-ai/go-tokenizers/tokenizers"
"github.com/yalue/onnxruntime_go"
)
// Embedder defines the interface for embedding text
type Embedder interface {
Embed(text string) ([]float32, error)
EmbedSlice(lines []string) ([][]float32, error)
}
// APIEmbedder implements embedder using an API (like Hugging Face, OpenAI, etc.)
type APIEmbedder struct {
logger *slog.Logger
client *http.Client
cfg *config.Config
}
func NewAPIEmbedder(l *slog.Logger, cfg *config.Config) *APIEmbedder {
return &APIEmbedder{
logger: l,
client: &http.Client{},
cfg: cfg,
}
}
func (a *APIEmbedder) Embed(text string) ([]float32, error) {
payload, err := json.Marshal(
map[string]any{"input": text, "encoding_format": "float"},
)
if err != nil {
a.logger.Error("failed to marshal payload", "err", err.Error())
return nil, err
}
req, err := http.NewRequest("POST", a.cfg.EmbedURL, bytes.NewReader(payload))
if err != nil {
a.logger.Error("failed to create new req", "err", err.Error())
return nil, err
}
if a.cfg.HFToken != "" {
req.Header.Add("Authorization", "Bearer "+a.cfg.HFToken)
}
resp, err := a.client.Do(req)
if err != nil {
a.logger.Error("failed to embed text", "err", err.Error())
return nil, err
}
defer resp.Body.Close()
if resp.StatusCode != 200 {
err = fmt.Errorf("non 200 response; code: %v", resp.StatusCode)
a.logger.Error(err.Error())
return nil, err
}
embResp := &models.LCPEmbedResp{}
if err := json.NewDecoder(resp.Body).Decode(&embResp); err != nil {
a.logger.Error("failed to decode embedding response", "err", err.Error())
return nil, err
}
if len(embResp.Data) == 0 || len(embResp.Data[0].Embedding) == 0 {
err = errors.New("empty embedding response")
a.logger.Error("empty embedding response")
return nil, err
}
return embResp.Data[0].Embedding, nil
}
func (a *APIEmbedder) EmbedSlice(lines []string) ([][]float32, error) {
payload, err := json.Marshal(
map[string]any{"input": lines, "encoding_format": "float"},
)
if err != nil {
a.logger.Error("failed to marshal payload", "err", err.Error())
return nil, err
}
req, err := http.NewRequest("POST", a.cfg.EmbedURL, bytes.NewReader(payload))
if err != nil {
a.logger.Error("failed to create new req", "err", err.Error())
return nil, err
}
if a.cfg.HFToken != "" {
req.Header.Add("Authorization", "Bearer "+a.cfg.HFToken)
}
resp, err := a.client.Do(req)
if err != nil {
a.logger.Error("failed to embed text", "err", err.Error())
return nil, err
}
defer resp.Body.Close()
if resp.StatusCode != 200 {
err = fmt.Errorf("non 200 response; code: %v", resp.StatusCode)
a.logger.Error(err.Error())
return nil, err
}
embResp := &models.LCPEmbedResp{}
if err := json.NewDecoder(resp.Body).Decode(&embResp); err != nil {
a.logger.Error("failed to decode embedding response", "err", err.Error())
return nil, err
}
if len(embResp.Data) == 0 {
err = errors.New("empty embedding response")
a.logger.Error("empty embedding response")
return nil, err
}
// Collect all embeddings from the response
embeddings := make([][]float32, len(embResp.Data))
for i := range embResp.Data {
if len(embResp.Data[i].Embedding) == 0 {
err = fmt.Errorf("empty embedding at index %d", i)
a.logger.Error("empty embedding", "index", i)
return nil, err
}
embeddings[i] = embResp.Data[i].Embedding
}
// Sort embeddings by index to match the order of input lines
// API responses may not be in order
for _, data := range embResp.Data {
if data.Index >= len(embeddings) || data.Index < 0 {
err = fmt.Errorf("invalid embedding index %d", data.Index)
a.logger.Error("invalid embedding index", "index", data.Index)
return nil, err
}
embeddings[data.Index] = data.Embedding
}
return embeddings, nil
}
// 1. Loading ONNX models locally
// 2. Using a Go ONNX runtime (like gorgonia/onnx or similar)
// 3. Converting text to embeddings without external API calls
type ONNXEmbedder struct {
session *onnxruntime_go.DynamicAdvancedSession
tokenizer *tokenizers.Tokenizer
dims int // 768, 512, 256, or 128 for Matryoshka
}
func (e *ONNXEmbedder) EmbedSlice(texts []string) ([][]float32, error) {
// Batch processing
inputs := e.prepareBatch(texts)
outputs := make([][]float32, len(texts))
// Run batch inference (much faster)
err := e.session.Run(inputs, outputs)
return outputs, err
}
func NewONNXEmbedder(modelPath string) (*ONNXEmbedder, error) {
// Load ONNX model
session, err := onnxruntime_go.NewDynamicAdvancedSession(
modelPath, // onnx/embedgemma/model_q4.onnx
[]string{"input_ids", "attention_mask"},
[]string{"sentence_embedding"},
nil,
)
if err != nil {
return nil, err
}
// Load tokenizer (from Hugging Face)
tokenizer, err := tokenizers.FromFile("./tokenizer.json")
return &ONNXEmbedder{
session: session,
tokenizer: tokenizer,
}, nil
}
func (e *ONNXEmbedder) Embed(text string) ([]float32, error) {
// Tokenize
tokens := e.tokenizer.Encode(text, true)
// Prepare inputs
inputIDs := []int64{tokens.GetIds()}
attentionMask := []int64{tokens.GetAttentionMask()}
// Run inference
output := onnxruntime_go.NewEmptyTensor[float32](
onnxruntime_go.NewShape(1, 768),
)
err := e.session.Run(
map[string]any{
"input_ids": inputIDs,
"attention_mask": attentionMask,
},
[]string{"sentence_embedding"},
[]any{&output},
)
return output.GetData(), nil
}
|
