Fix bug

Author: xusenlin

README.md

@@ -20,7 +20,7 @@
 
 ## 📢 新闻
 
-+ 【2023.11.09】 `dev` 分支已经支持 `openai=1.2.0` 版本
++ 【2023.11.09】 目前分支需要安装 `openai>=1.2.3` 版本
 
 
 + 【2023.11.03】 支持 `chatglm3` 和 `qwen` 模型的 `function call` 调用功能，同时支持流式和非流式模式, [工具使用示例](https://github.com/xusenlinzy/api-for-open-llm/tree/master/examples/chatglm3/tool_using.py), 网页 `demo` 已经集成到 [streamlit-demo](./streamlit-demo)

api/generation/chatglm.py

@@ -82,7 +82,7 @@ def generate_stream_chatglm(
     context_len=2048,
     stream_interval=2,
 ):
-    prompt = params["messages"]
+    prompt = params["prompt"]
     temperature = float(params.get("temperature", 1.0))
     repetition_penalty = float(params.get("repetition_penalty", 1.0))
     top_p = float(params.get("top_p", 1.0))

api/routes/embedding.py

@@ -1,13 +1,14 @@
+import base64
+
 import numpy as np
 import tiktoken
 from fastapi import APIRouter, Depends
-from openai.types.create_embedding_response import CreateEmbeddingResponse, Usage
-from openai.types.embedding import Embedding
+from openai.types.create_embedding_response import Usage
 
 from api.config import config
 from api.models import EMBEDDED_MODEL
 from api.routes.utils import check_api_key
-from api.utils.protocol import EmbeddingCreateParams
+from api.utils.protocol import EmbeddingCreateParams, Embedding, CreateEmbeddingResponse
 
 embedding_router = APIRouter()
 
@@ -19,16 +20,16 @@ async def create_embeddings(request: EmbeddingCreateParams, model_name: str = No
     if request.model is None:
         request.model = model_name
 
-    inputs = request.input
-    if isinstance(inputs, str):
-        inputs = [inputs]
-    elif isinstance(inputs, list):
-        if isinstance(inputs[0], int):
+    request.input = request.input
+    if isinstance(request.input, str):
+        request.input = [request.input]
+    elif isinstance(request.input, list):
+        if isinstance(request.input[0], int):
             decoding = tiktoken.model.encoding_for_model(request.model)
-            inputs = [decoding.decode(inputs)]
-        elif isinstance(inputs[0], list):
+            request.input = [decoding.decode(request.input)]
+        elif isinstance(request.input[0], list):
             decoding = tiktoken.model.encoding_for_model(request.model)
-            inputs = [decoding.decode(text) for text in inputs]
+            request.input = [decoding.decode(text) for text in request.input]
 
     # https://huggingface.co/BAAI/bge-large-zh
     if EMBEDDED_MODEL is not None:
@@ -38,12 +39,11 @@ async def create_embeddings(request: EmbeddingCreateParams, model_name: str = No
                 instruction = "为这个句子生成表示以用于检索相关文章："
             elif "en" in config.EMBEDDING_NAME.lower():
                 instruction = "Represent this sentence for searching relevant passages: "
-            inputs = [instruction + q for q in inputs]
+            request.inputs = [instruction + q for q in request.input]
 
     data, total_tokens = [], 0
     batches = [
-        inputs[i: min(i + 1024, len(inputs))]
-        for i in range(0, len(inputs), 1024)
+        request.input[i: i + 1024] for i in range(0, len(request.input), 1024)
     ]
     for num_batch, batch in enumerate(batches):
         token_num = sum([len(i) for i in batch])
@@ -54,10 +54,14 @@ async def create_embeddings(request: EmbeddingCreateParams, model_name: str = No
             zeros = np.zeros((bs, config.EMBEDDING_SIZE - dim))
             vecs = np.c_[vecs, zeros]
 
-        vecs = vecs.tolist()
+        if request.encoding_format == "base64":
+            vecs = [base64.b64encode(v.tobytes()).decode("utf-8") for v in vecs]
+        else:
+            vecs = vecs.tolist()
+
         for i, embed in enumerate(vecs):
             data.append(
-                Embedding(index=num_batch * 1024 + i, embedding=embed, object="embedding")
+                Embedding(index=num_batch * 1024 + i, object="embedding", embedding=embed)
             )
 
         total_tokens += token_num

api/utils/protocol.py

@@ -1,12 +1,13 @@
 from enum import Enum
-from typing import Optional, Dict, List, Union, Literal
+from typing import Optional, Dict, List, Union, Literal, Any
 
 from openai.types.chat import (
     ChatCompletionMessageParam,
     ChatCompletionToolChoiceOptionParam,
     ChatCompletionToolParam,
 )
 from openai.types.chat.completion_create_params import FunctionCall, ResponseFormat
+from openai.types.create_embedding_response import Usage
 from pydantic import BaseModel
 
 
@@ -358,3 +359,32 @@ class EmbeddingCreateParams(BaseModel):
     and detect abuse.
     [Learn more](https://platform.openai.com/docs/guides/safety-best-practices/end-user-ids).
     """
+
+
+class Embedding(BaseModel):
+    embedding: Any
+    """The embedding vector, which is a list of floats.
+
+    The length of vector depends on the model as listed in the
+    [embedding guide](https://platform.openai.com/docs/guides/embeddings).
+    """
+
+    index: int
+    """The index of the embedding in the list of embeddings."""
+
+    object: Literal["embedding"]
+    """The object type, which is always "embedding"."""
+
+
+class CreateEmbeddingResponse(BaseModel):
+    data: List[Embedding]
+    """The list of embeddings generated by the model."""
+
+    model: str
+    """The name of the model used to generate the embedding."""
+
+    object: Literal["list"]
+    """The object type, which is always "list"."""
+
+    usage: Usage
+    """The usage information for the request."""

requirements.txt

@@ -1,4 +1,4 @@
-openai==1.2.0
+openai>=1.2.3
 bitsandbytes
 fastapi==0.95.1
 typing-inspect==0.8.0

tests/langchain_test.py

@@ -0,0 +1,13 @@
+from langchain.chat_models import ChatOpenAI
+from langchain.embeddings import OpenAIEmbeddings
+from langchain.schema import HumanMessage
+
+text = "你好"
+messages = [HumanMessage(content=text)]
+
+llm = ChatOpenAI(openai_api_key="xxx", openai_api_base="http://192.168.20.59:7891/v1")
+
+print(llm(messages))
+
+embedding = OpenAIEmbeddings(openai_api_key="xxx", openai_api_base="http://192.168.20.59:7891/v1")
+print(embedding.embed_documents(["你好"]))