471 lines
16 KiB
Go
471 lines
16 KiB
Go
/*
|
|
* Copyright 2025 coze-dev Authors
|
|
*
|
|
* 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 knowledge
|
|
|
|
import (
|
|
"context"
|
|
"encoding/json"
|
|
"fmt"
|
|
"os"
|
|
"path/filepath"
|
|
"strconv"
|
|
"time"
|
|
|
|
"github.com/cloudwego/eino-ext/components/embedding/ark"
|
|
ollamaEmb "github.com/cloudwego/eino-ext/components/embedding/ollama"
|
|
"github.com/cloudwego/eino-ext/components/embedding/openai"
|
|
ao "github.com/cloudwego/eino-ext/components/model/ark"
|
|
"github.com/cloudwego/eino-ext/components/model/deepseek"
|
|
"github.com/cloudwego/eino-ext/components/model/gemini"
|
|
"github.com/cloudwego/eino-ext/components/model/ollama"
|
|
mo "github.com/cloudwego/eino-ext/components/model/openai"
|
|
"github.com/cloudwego/eino-ext/components/model/qwen"
|
|
"github.com/cloudwego/eino/components/prompt"
|
|
"github.com/cloudwego/eino/schema"
|
|
"github.com/milvus-io/milvus/client/v2/milvusclient"
|
|
"github.com/volcengine/volc-sdk-golang/service/vikingdb"
|
|
"github.com/volcengine/volc-sdk-golang/service/visual"
|
|
"google.golang.org/genai"
|
|
"gorm.io/gorm"
|
|
|
|
"github.com/coze-dev/coze-studio/backend/application/search"
|
|
knowledgeImpl "github.com/coze-dev/coze-studio/backend/domain/knowledge/service"
|
|
"github.com/coze-dev/coze-studio/backend/infra/contract/cache"
|
|
"github.com/coze-dev/coze-studio/backend/infra/contract/chatmodel"
|
|
"github.com/coze-dev/coze-studio/backend/infra/contract/document/nl2sql"
|
|
"github.com/coze-dev/coze-studio/backend/infra/contract/document/ocr"
|
|
"github.com/coze-dev/coze-studio/backend/infra/contract/document/searchstore"
|
|
"github.com/coze-dev/coze-studio/backend/infra/contract/embedding"
|
|
"github.com/coze-dev/coze-studio/backend/infra/contract/es"
|
|
"github.com/coze-dev/coze-studio/backend/infra/contract/idgen"
|
|
"github.com/coze-dev/coze-studio/backend/infra/contract/imagex"
|
|
"github.com/coze-dev/coze-studio/backend/infra/contract/messages2query"
|
|
"github.com/coze-dev/coze-studio/backend/infra/contract/rdb"
|
|
"github.com/coze-dev/coze-studio/backend/infra/contract/storage"
|
|
chatmodelImpl "github.com/coze-dev/coze-studio/backend/infra/impl/chatmodel"
|
|
builtinNL2SQL "github.com/coze-dev/coze-studio/backend/infra/impl/document/nl2sql/builtin"
|
|
"github.com/coze-dev/coze-studio/backend/infra/impl/document/ocr/veocr"
|
|
builtinParser "github.com/coze-dev/coze-studio/backend/infra/impl/document/parser/builtin"
|
|
"github.com/coze-dev/coze-studio/backend/infra/impl/document/rerank/rrf"
|
|
sses "github.com/coze-dev/coze-studio/backend/infra/impl/document/searchstore/elasticsearch"
|
|
ssmilvus "github.com/coze-dev/coze-studio/backend/infra/impl/document/searchstore/milvus"
|
|
ssvikingdb "github.com/coze-dev/coze-studio/backend/infra/impl/document/searchstore/vikingdb"
|
|
arkemb "github.com/coze-dev/coze-studio/backend/infra/impl/embedding/ark"
|
|
"github.com/coze-dev/coze-studio/backend/infra/impl/embedding/wrap"
|
|
"github.com/coze-dev/coze-studio/backend/infra/impl/eventbus"
|
|
builtinM2Q "github.com/coze-dev/coze-studio/backend/infra/impl/messages2query/builtin"
|
|
"github.com/coze-dev/coze-studio/backend/pkg/lang/conv"
|
|
"github.com/coze-dev/coze-studio/backend/pkg/lang/ptr"
|
|
"github.com/coze-dev/coze-studio/backend/pkg/logs"
|
|
"github.com/coze-dev/coze-studio/backend/types/consts"
|
|
)
|
|
|
|
type ServiceComponents struct {
|
|
DB *gorm.DB
|
|
IDGenSVC idgen.IDGenerator
|
|
Storage storage.Storage
|
|
RDB rdb.RDB
|
|
ImageX imagex.ImageX
|
|
ES es.Client
|
|
EventBus search.ResourceEventBus
|
|
CacheCli cache.Cmdable
|
|
}
|
|
|
|
func InitService(c *ServiceComponents) (*KnowledgeApplicationService, error) {
|
|
ctx := context.Background()
|
|
|
|
nameServer := os.Getenv(consts.MQServer)
|
|
|
|
knowledgeProducer, err := eventbus.NewProducer(nameServer, consts.RMQTopicKnowledge, consts.RMQConsumeGroupKnowledge, 2)
|
|
if err != nil {
|
|
return nil, fmt.Errorf("init knowledge producer failed, err=%w", err)
|
|
}
|
|
|
|
var sManagers []searchstore.Manager
|
|
|
|
// es full text search
|
|
sManagers = append(sManagers, sses.NewManager(&sses.ManagerConfig{Client: c.ES}))
|
|
|
|
// vector search
|
|
mgr, err := getVectorStore(ctx)
|
|
if err != nil {
|
|
return nil, fmt.Errorf("init vector store failed, err=%w", err)
|
|
}
|
|
sManagers = append(sManagers, mgr)
|
|
|
|
var ocrImpl ocr.OCR
|
|
switch os.Getenv("OCR_TYPE") {
|
|
case "ve":
|
|
ocrAK := os.Getenv("VE_OCR_AK")
|
|
ocrSK := os.Getenv("VE_OCR_SK")
|
|
if ocrAK == "" || ocrSK == "" {
|
|
logs.Warnf("[ve_ocr] ak / sk not configured, ocr might not work well")
|
|
}
|
|
inst := visual.NewInstance()
|
|
inst.Client.SetAccessKey(ocrAK)
|
|
inst.Client.SetSecretKey(ocrSK)
|
|
ocrImpl = veocr.NewOCR(&veocr.Config{Client: inst})
|
|
default:
|
|
// accept ocr not configured
|
|
}
|
|
|
|
root, err := os.Getwd()
|
|
if err != nil {
|
|
logs.Warnf("[InitConfig] Failed to get current working directory: %v", err)
|
|
root = os.Getenv("PWD")
|
|
}
|
|
|
|
var rewriter messages2query.MessagesToQuery
|
|
if rewriterChatModel, _, err := getBuiltinChatModel(ctx, "M2Q_"); err != nil {
|
|
return nil, err
|
|
} else {
|
|
filePath := filepath.Join(root, "resources/conf/prompt/messages_to_query_template_jinja2.json")
|
|
rewriterTemplate, err := readJinja2PromptTemplate(filePath)
|
|
if err != nil {
|
|
return nil, err
|
|
}
|
|
rewriter, err = builtinM2Q.NewMessagesToQuery(ctx, rewriterChatModel, rewriterTemplate)
|
|
if err != nil {
|
|
return nil, err
|
|
}
|
|
}
|
|
|
|
var n2s nl2sql.NL2SQL
|
|
if n2sChatModel, _, err := getBuiltinChatModel(ctx, "NL2SQL_"); err != nil {
|
|
return nil, err
|
|
} else {
|
|
filePath := filepath.Join(root, "resources/conf/prompt/nl2sql_template_jinja2.json")
|
|
n2sTemplate, err := readJinja2PromptTemplate(filePath)
|
|
if err != nil {
|
|
return nil, err
|
|
}
|
|
n2s, err = builtinNL2SQL.NewNL2SQL(ctx, n2sChatModel, n2sTemplate)
|
|
if err != nil {
|
|
return nil, err
|
|
}
|
|
}
|
|
|
|
imageAnnoChatModel, configured, err := getBuiltinChatModel(ctx, "IA_")
|
|
if err != nil {
|
|
return nil, err
|
|
}
|
|
|
|
knowledgeDomainSVC, knowledgeEventHandler := knowledgeImpl.NewKnowledgeSVC(&knowledgeImpl.KnowledgeSVCConfig{
|
|
DB: c.DB,
|
|
IDGen: c.IDGenSVC,
|
|
RDB: c.RDB,
|
|
Producer: knowledgeProducer,
|
|
SearchStoreManagers: sManagers,
|
|
ParseManager: builtinParser.NewManager(c.Storage, ocrImpl, imageAnnoChatModel), // default builtin
|
|
Storage: c.Storage,
|
|
Rewriter: rewriter,
|
|
Reranker: rrf.NewRRFReranker(0), // default rrf
|
|
NL2Sql: n2s,
|
|
OCR: ocrImpl,
|
|
CacheCli: c.CacheCli,
|
|
IsAutoAnnotationSupported: configured,
|
|
ModelFactory: chatmodelImpl.NewDefaultFactory(),
|
|
})
|
|
|
|
if err = eventbus.RegisterConsumer(nameServer, consts.RMQTopicKnowledge, consts.RMQConsumeGroupKnowledge, knowledgeEventHandler); err != nil {
|
|
return nil, fmt.Errorf("register knowledge consumer failed, err=%w", err)
|
|
}
|
|
|
|
KnowledgeSVC.DomainSVC = knowledgeDomainSVC
|
|
KnowledgeSVC.eventBus = c.EventBus
|
|
KnowledgeSVC.storage = c.Storage
|
|
return KnowledgeSVC, nil
|
|
}
|
|
|
|
func getVectorStore(ctx context.Context) (searchstore.Manager, error) {
|
|
vsType := os.Getenv("VECTOR_STORE_TYPE")
|
|
|
|
switch vsType {
|
|
case "milvus":
|
|
cctx, cancel := context.WithTimeout(ctx, time.Second*5)
|
|
defer cancel()
|
|
|
|
milvusAddr := os.Getenv("MILVUS_ADDR")
|
|
mc, err := milvusclient.New(cctx, &milvusclient.ClientConfig{Address: milvusAddr})
|
|
if err != nil {
|
|
return nil, fmt.Errorf("init milvus client failed, err=%w", err)
|
|
}
|
|
|
|
emb, err := getEmbedding(ctx)
|
|
if err != nil {
|
|
return nil, fmt.Errorf("init milvus embedding failed, err=%w", err)
|
|
}
|
|
|
|
mgr, err := ssmilvus.NewManager(&ssmilvus.ManagerConfig{
|
|
Client: mc,
|
|
Embedding: emb,
|
|
EnableHybrid: ptr.Of(true),
|
|
})
|
|
if err != nil {
|
|
return nil, fmt.Errorf("init milvus vector store failed, err=%w", err)
|
|
}
|
|
|
|
return mgr, nil
|
|
case "vikingdb":
|
|
var (
|
|
host = os.Getenv("VIKING_DB_HOST")
|
|
region = os.Getenv("VIKING_DB_REGION")
|
|
ak = os.Getenv("VIKING_DB_AK")
|
|
sk = os.Getenv("VIKING_DB_SK")
|
|
scheme = os.Getenv("VIKING_DB_SCHEME")
|
|
modelName = os.Getenv("VIKING_DB_MODEL_NAME")
|
|
)
|
|
if ak == "" || sk == "" {
|
|
return nil, fmt.Errorf("invalid vikingdb ak / sk")
|
|
}
|
|
if host == "" {
|
|
host = "api-vikingdb.volces.com"
|
|
}
|
|
if region == "" {
|
|
region = "cn-beijing"
|
|
}
|
|
if scheme == "" {
|
|
scheme = "https"
|
|
}
|
|
|
|
var embConfig *ssvikingdb.VikingEmbeddingConfig
|
|
if modelName != "" {
|
|
embName := ssvikingdb.VikingEmbeddingModelName(modelName)
|
|
if embName.Dimensions() == 0 {
|
|
return nil, fmt.Errorf("embedding model not support, model_name=%s", modelName)
|
|
}
|
|
embConfig = &ssvikingdb.VikingEmbeddingConfig{
|
|
UseVikingEmbedding: true,
|
|
EnableHybrid: embName.SupportStatus() == embedding.SupportDenseAndSparse,
|
|
ModelName: embName,
|
|
ModelVersion: embName.ModelVersion(),
|
|
DenseWeight: ptr.Of(0.2),
|
|
BuiltinEmbedding: nil,
|
|
}
|
|
} else {
|
|
builtinEmbedding, err := getEmbedding(ctx)
|
|
if err != nil {
|
|
return nil, fmt.Errorf("builtint embedding init failed, err=%w", err)
|
|
}
|
|
|
|
embConfig = &ssvikingdb.VikingEmbeddingConfig{
|
|
UseVikingEmbedding: false,
|
|
EnableHybrid: false,
|
|
BuiltinEmbedding: builtinEmbedding,
|
|
}
|
|
}
|
|
svc := vikingdb.NewVikingDBService(host, region, ak, sk, scheme)
|
|
mgr, err := ssvikingdb.NewManager(&ssvikingdb.ManagerConfig{
|
|
Service: svc,
|
|
IndexingConfig: nil, // use default config
|
|
EmbeddingConfig: embConfig,
|
|
})
|
|
if err != nil {
|
|
return nil, fmt.Errorf("init vikingdb manager failed, err=%w", err)
|
|
}
|
|
|
|
return mgr, nil
|
|
|
|
default:
|
|
return nil, fmt.Errorf("unexpected vector store type, type=%s", vsType)
|
|
}
|
|
}
|
|
|
|
func getEmbedding(ctx context.Context) (embedding.Embedder, error) {
|
|
var emb embedding.Embedder
|
|
|
|
switch os.Getenv("EMBEDDING_TYPE") {
|
|
case "openai":
|
|
var (
|
|
openAIEmbeddingBaseURL = os.Getenv("OPENAI_EMBEDDING_BASE_URL")
|
|
openAIEmbeddingModel = os.Getenv("OPENAI_EMBEDDING_MODEL")
|
|
openAIEmbeddingApiKey = os.Getenv("OPENAI_EMBEDDING_API_KEY")
|
|
openAIEmbeddingByAzure = os.Getenv("OPENAI_EMBEDDING_BY_AZURE")
|
|
openAIEmbeddingApiVersion = os.Getenv("OPENAI_EMBEDDING_API_VERSION")
|
|
openAIEmbeddingDims = os.Getenv("OPENAI_EMBEDDING_DIMS")
|
|
openAIRequestEmbeddingDims = os.Getenv("OPENAI_EMBEDDING_REQUEST_DIMS")
|
|
)
|
|
|
|
byAzure, err := strconv.ParseBool(openAIEmbeddingByAzure)
|
|
if err != nil {
|
|
return nil, fmt.Errorf("init openai embedding by_azure failed, err=%w", err)
|
|
}
|
|
|
|
dims, err := strconv.ParseInt(openAIEmbeddingDims, 10, 64)
|
|
if err != nil {
|
|
return nil, fmt.Errorf("init openai embedding dims failed, err=%w", err)
|
|
}
|
|
|
|
openAICfg := &openai.EmbeddingConfig{
|
|
APIKey: openAIEmbeddingApiKey,
|
|
ByAzure: byAzure,
|
|
BaseURL: openAIEmbeddingBaseURL,
|
|
APIVersion: openAIEmbeddingApiVersion,
|
|
Model: openAIEmbeddingModel,
|
|
// Dimensions: ptr.Of(int(dims)),
|
|
}
|
|
reqDims := conv.StrToInt64D(openAIRequestEmbeddingDims, 0)
|
|
if reqDims > 0 {
|
|
// some openai model not support request dims
|
|
openAICfg.Dimensions = ptr.Of(int(reqDims))
|
|
}
|
|
|
|
emb, err = wrap.NewOpenAIEmbedder(ctx, openAICfg, dims)
|
|
if err != nil {
|
|
return nil, fmt.Errorf("init openai embedding failed, err=%w", err)
|
|
}
|
|
|
|
case "ark":
|
|
var (
|
|
arkEmbeddingBaseURL = os.Getenv("ARK_EMBEDDING_BASE_URL")
|
|
arkEmbeddingModel = os.Getenv("ARK_EMBEDDING_MODEL")
|
|
arkEmbeddingAK = os.Getenv("ARK_EMBEDDING_AK")
|
|
arkEmbeddingDims = os.Getenv("ARK_EMBEDDING_DIMS")
|
|
)
|
|
|
|
dims, err := strconv.ParseInt(arkEmbeddingDims, 10, 64)
|
|
if err != nil {
|
|
return nil, fmt.Errorf("init ark embedding dims failed, err=%w", err)
|
|
}
|
|
|
|
emb, err = arkemb.NewArkEmbedder(ctx, &ark.EmbeddingConfig{
|
|
APIKey: arkEmbeddingAK,
|
|
Model: arkEmbeddingModel,
|
|
BaseURL: arkEmbeddingBaseURL,
|
|
}, dims)
|
|
if err != nil {
|
|
return nil, fmt.Errorf("init ark embedding client failed, err=%w", err)
|
|
}
|
|
|
|
case "ollama":
|
|
var (
|
|
ollamaEmbeddingBaseURL = os.Getenv("OLLAMA_EMBEDDING_BASE_URL")
|
|
ollamaEmbeddingModel = os.Getenv("OLLAMA_EMBEDDING_MODEL")
|
|
ollamaEmbeddingDims = os.Getenv("OLLAMA_EMBEDDING_DIMS")
|
|
)
|
|
|
|
dims, err := strconv.ParseInt(ollamaEmbeddingDims, 10, 64)
|
|
if err != nil {
|
|
return nil, fmt.Errorf("init ollama embedding dims failed, err=%w", err)
|
|
}
|
|
|
|
emb, err = wrap.NewOllamaEmbedder(ctx, &ollamaEmb.EmbeddingConfig{
|
|
BaseURL: ollamaEmbeddingBaseURL,
|
|
Model: ollamaEmbeddingModel,
|
|
}, dims)
|
|
if err != nil {
|
|
return nil, fmt.Errorf("init ollama embedding failed, err=%w", err)
|
|
}
|
|
|
|
default:
|
|
return nil, fmt.Errorf("init knowledge embedding failed, type not configured")
|
|
}
|
|
|
|
return emb, nil
|
|
}
|
|
|
|
func getBuiltinChatModel(ctx context.Context, envPrefix string) (bcm chatmodel.BaseChatModel, configured bool, err error) {
|
|
getEnv := func(key string) string {
|
|
if val := os.Getenv(envPrefix + key); val != "" {
|
|
return val
|
|
}
|
|
return os.Getenv(key)
|
|
}
|
|
|
|
switch getEnv("BUILTIN_CM_TYPE") {
|
|
case "openai":
|
|
byAzure, _ := strconv.ParseBool(getEnv("BUILTIN_CM_OPENAI_BY_AZURE"))
|
|
bcm, err = mo.NewChatModel(ctx, &mo.ChatModelConfig{
|
|
APIKey: getEnv("BUILTIN_CM_OPENAI_API_KEY"),
|
|
ByAzure: byAzure,
|
|
BaseURL: getEnv("BUILTIN_CM_OPENAI_BASE_URL"),
|
|
Model: getEnv("BUILTIN_CM_OPENAI_MODEL"),
|
|
})
|
|
case "ark":
|
|
bcm, err = ao.NewChatModel(ctx, &ao.ChatModelConfig{
|
|
APIKey: getEnv("BUILTIN_CM_ARK_API_KEY"),
|
|
Model: getEnv("BUILTIN_CM_ARK_MODEL"),
|
|
BaseURL: getEnv("BUILTIN_CM_ARK_BASE_URL"),
|
|
})
|
|
case "deepseek":
|
|
bcm, err = deepseek.NewChatModel(ctx, &deepseek.ChatModelConfig{
|
|
APIKey: getEnv("BUILTIN_CM_DEEPSEEK_API_KEY"),
|
|
BaseURL: getEnv("BUILTIN_CM_DEEPSEEK_BASE_URL"),
|
|
Model: getEnv("BUILTIN_CM_DEEPSEEK_MODEL"),
|
|
})
|
|
case "ollama":
|
|
bcm, err = ollama.NewChatModel(ctx, &ollama.ChatModelConfig{
|
|
BaseURL: getEnv("BUILTIN_CM_OLLAMA_BASE_URL"),
|
|
Model: getEnv("BUILTIN_CM_OLLAMA_MODEL"),
|
|
})
|
|
case "qwen":
|
|
bcm, err = qwen.NewChatModel(ctx, &qwen.ChatModelConfig{
|
|
APIKey: getEnv("BUILTIN_CM_QWEN_API_KEY"),
|
|
BaseURL: getEnv("BUILTIN_CM_QWEN_BASE_URL"),
|
|
Model: getEnv("BUILTIN_CM_QWEN_MODEL"),
|
|
})
|
|
case "gemini":
|
|
backend, convErr := strconv.ParseInt(getEnv("BUILTIN_CM_GEMINI_BACKEND"), 10, 64)
|
|
if convErr != nil {
|
|
return nil, false, convErr
|
|
}
|
|
c, clientErr := genai.NewClient(ctx, &genai.ClientConfig{
|
|
APIKey: getEnv("BUILTIN_CM_GEMINI_API_KEY"),
|
|
Backend: genai.Backend(backend),
|
|
Project: getEnv("BUILTIN_CM_GEMINI_PROJECT"),
|
|
Location: getEnv("BUILTIN_CM_GEMINI_LOCATION"),
|
|
HTTPOptions: genai.HTTPOptions{
|
|
BaseURL: getEnv("BUILTIN_CM_GEMINI_BASE_URL"),
|
|
},
|
|
})
|
|
if clientErr != nil {
|
|
return nil, false, clientErr
|
|
}
|
|
bcm, err = gemini.NewChatModel(ctx, &gemini.Config{
|
|
Client: c,
|
|
Model: getEnv("BUILTIN_CM_GEMINI_MODEL"),
|
|
})
|
|
default:
|
|
// accept builtin chat model not configured
|
|
}
|
|
|
|
if err != nil {
|
|
return nil, false, fmt.Errorf("knowledge init openai chat mode failed, %w", err)
|
|
}
|
|
if bcm != nil {
|
|
configured = true
|
|
}
|
|
|
|
return
|
|
}
|
|
|
|
func readJinja2PromptTemplate(jsonFilePath string) (prompt.ChatTemplate, error) {
|
|
b, err := os.ReadFile(jsonFilePath)
|
|
if err != nil {
|
|
return nil, err
|
|
}
|
|
var m2qMessages []*schema.Message
|
|
if err = json.Unmarshal(b, &m2qMessages); err != nil {
|
|
return nil, err
|
|
}
|
|
tpl := make([]schema.MessagesTemplate, len(m2qMessages))
|
|
for i := range m2qMessages {
|
|
tpl[i] = m2qMessages[i]
|
|
}
|
|
return prompt.FromMessages(schema.Jinja2, tpl...), nil
|
|
}
|