rag

.gitignore

@@ -1,6 +1,4 @@
 .DS_Store
-configs/*
-!configs/example.toml
 llm_backend/protos/
 
 # Byte-compiled / optimized / DLL files

README.md

@@ -11,7 +11,7 @@ uv run gen-protos
 
 ## Usage
 
-Please configure the `configs/config.toml` file (refer to `configs/example.toml` for the options).
+Please configure the `configs/config.toml` file.
 The following environment variables are required (`export` them or place them in a `.env` file):
 
 - `OPENAI_API_KEY`: Your ChatGPT API key.
@@ -20,7 +20,13 @@ The following environment variables are required (`export` them or place them in
 - `QDRANT_COLLECTION`: The Qdrant collection name.
 
 ```shell
-uv run serve --config configs/config.toml
+python3 scripts/serve.py --config configs/config.toml
+```
+
+You can refer to `scripts/client.py` for an example implementation of a client:
+
+```shell
+python3 scripts/client.py
 ```
 
 ## Features

configs/config.toml

@@ -0,0 +1,44 @@
+[server]
+host = 'localhost'
+port = 50051
+max_workers = 10
+
+[service.retrieve]
+# Name of embedding model. All available models can be found [here](https://huggingface.co/models?language=zh)
+embedding_model = 'intfloat/multilingual-e5-large'
+
+# The template must contain the `{keywords}` placeholder.
+prompt_template = 'Please search for the content related to the following keywords: {keywords}.'
+similarity_top_k = 5
+
+[service.summarize]
+system_template = """
+# Project Mission: My project mission is to extract 5 articles of the same type from the internet each time and provide them to Chat GPT in the same format to generate summaries and digests, making it convenient for the general public to read.
+# Input Format: The format during input is as follows: 1.xxx 2.xxx 3.xxx 4.xxx 5.xxx Each news article is numbered with a digit title. There is a blank line between different news articles, but within the same article, there are no line breaks.
+# Detailed Project Execution: The detailed execution of the project involves refraining from adding personal opinions. I only generate summaries based on the provided news and refrain from providing responses beyond the scope of the news.
+# Audience for My Content: The audience comprises professionals from various fields, as well as students and homemakers. They span a wide range of age groups and have a strong desire for knowledge. However, due to limited time, they cannot personally read a large amount of news information. Therefore, my content typically needs to be transformed into something understandable by the general public, with logical analysis involving specific questions and solutions.
+
+# Assuming you are now a reader, think step by step about what you think the key points of the news would be, and provide the first version of the summary. Then, based on this summary, pose sub-questions and further modify to provide the final summary.
+# Answer in Traditional Chinese, and refrain from providing thoughts and content beyond what you've provided. Endeavor to comprehensively describe the key points of the news.
+# Responses should strive to be rigorous and formal, with real evidence when answering questions.
+# Answers can be as complete and comprehensive as possible, expanding on details and actual content.
+# The "Output Format" is: provide an overarching title that summarizes the news content above, then summarizes the content.
+"""
+
+# The template must contain the `{context_str}` and `{query_str}` placeholders.
+user_template = """
+{query_str}
+---------------------
+{context_str}"""
+
+# The content of `{query_str}` placeholder in the user template.
+query_str = '假設你是一個摘要抓取者，請將以下---內的文字做一篇文章摘要，用文章敘述的方式呈現，不要用列點的，至少要有500字，要有標題。'
+
+# The transform function from the request strings to the query strings.
+# Must be one of:
+# - 'plain': The query string is the same as the request string.
+# - 'numbered': Add a number (1., 2., ...) to the beginning of each request string.
+content_format = 'plain'
+
+[service.summarize.llm]
+model = 'gpt-4o-mini'

llm_backend/__init__.py

@@ -1,18 +1,9 @@
 import os
 
-from grpc._server import _Server
+import grpc
 from pydantic import BaseModel
 
-from .search import (
-    SearchService,
-    add_SearchServiceServicer_to_server,
-)
-from .search.config import SearchConfig
-from .summarize import (
-    SummarizeService,
-    add_SummarizeServiceServicer_to_server,
-)
-from .summarize.config import SummarizeConfig
+from .rag import RagConfig, RagService, add_RagServiceServicer_to_server
 
 
 class ServerConfig(BaseModel):
@@ -21,21 +12,11 @@ class ServerConfig(BaseModel):
     max_workers: int = (os.cpu_count() or 1) * 5
 
 
-class ServiceConfig(BaseModel):
-    search: SearchConfig
-    summarize: SummarizeConfig
-
-
 class Config(BaseModel):
     server: ServerConfig
-    service: ServiceConfig
-
-
-def setup_search_service(config: Config, server: _Server):
-    search_service = SearchService(config.service.search)
-    add_SearchServiceServicer_to_server(search_service, server)
+    service: RagConfig
 
 
-def setup_summarize_service(config: Config, server: _Server):
-    summarize_service = SummarizeService(config.service.summarize)
-    add_SummarizeServiceServicer_to_server(summarize_service, server)
+def setup_rag_service(config: Config, server: grpc.aio.Server):
+    rag_service = RagService(config.service)
+    add_RagServiceServicer_to_server(rag_service, server)

llm_backend/rag/__init__.py

@@ -0,0 +1,5 @@
+from ..protos.rag_pb2_grpc import (
+    add_RagServiceServicer_to_server as add_RagServiceServicer_to_server,
+)
+from .config import RagConfig as RagConfig
+from .service import RagService as RagService

llm_backend/summarize/config.py → llm_backend/rag/config.py

@@ -1,12 +1,17 @@
-from enum import StrEnum
 from typing import Annotated
 
 from llama_index.llms.openai.utils import ALL_AVAILABLE_MODELS
 from pydantic import AfterValidator, BaseModel, Field
-from pydantic_settings import BaseSettings, SettingsConfigDict
+from pydantic_settings import BaseSettings
 
-from ..utils import contains_placeholder
+from .content_formatters import ContentFormat
 
+DEFAULT_EMBEDDING_MODEL = "intfloat/multilingual-e5-large"
+DEFAULT_OPENAI_MODEL = "gpt-4o-mini"
+DEFAULT_SIMILARITY_TOP_K = 10
+DEFAULT_QUERY_PROMPT_TEMPLATE = (
+    "Please search for the content related to the following keywords: {keywords}."
+)
 DEFAULT_SYSTEM_TEMPLATE = (
     "You are an expert Q&A system that is trusted around the world.\n"
     "Always answer the query using the provided context information,"
@@ -30,9 +35,33 @@
 DEFAULT_QUERY_STR = "請用繁體中文總結這幾篇新聞。"
 
 
-class ContentFormat(StrEnum):
-    PLAIN = "plain"
-    NUMBERED = "numbered"
+def contains_placeholder(*placeholders: str):
+    def validate_template(template: str):
+        for placeholder in placeholders:
+            if f"{{{placeholder}}}" not in template:
+                raise ValueError(f"Template must contain '{{{placeholder}}}'")
+        return template
+
+    return validate_template
+
+
+class QDrantConfig(BaseSettings):
+    host: str = Field("test", validation_alias="QDRANT_HOST")
+    port: int = Field(6333, gt=0, validation_alias="QDRANT_PORT")
+    collection: str = Field("news", validation_alias="QDRANT_COLLECTION")
+
+
+class RetrieveConfig(BaseModel):
+    vector_database: QDrantConfig = QDrantConfig()  # type: ignore
+    embedding_model: str = Field(
+        DEFAULT_EMBEDDING_MODEL,
+        description="Name of embedding model."
+        "All available models can be found [here](https://huggingface.co/models?library=sentence-transformers&language=zh).",
+    )
+    prompt_template: Annotated[
+        str, AfterValidator(contains_placeholder("keywords"))
+    ] = DEFAULT_QUERY_PROMPT_TEMPLATE
+    similarity_top_k: int = Field(DEFAULT_SIMILARITY_TOP_K, gt=1)
 
 
 def is_available_model(model_name: str):
@@ -43,23 +72,17 @@ def is_available_model(model_name: str):
     return model_name
 
 
-class ChatgptConfig(BaseSettings):
-    model_config = SettingsConfigDict(
-        env_file=(".env", ".env.prod"),
-        env_file_encoding="utf-8",
-        case_sensitive=True,
-        extra="ignore",
-    )
-
+class ChatGptConfig(BaseSettings):
     api_key: str = Field(validation_alias="OPENAI_API_KEY")
     model: Annotated[
         str,
-        Field("gpt-3.5-turbo"),
+        Field(DEFAULT_OPENAI_MODEL),
         AfterValidator(is_available_model),
     ]
 
 
-class SummarizeQueryConfig(BaseModel):
+class SummarizeConfig(BaseModel):
+    llm: ChatGptConfig = ChatGptConfig()  # type: ignore
     system_template: str = DEFAULT_SYSTEM_TEMPLATE
     user_template: Annotated[
         str, AfterValidator(contains_placeholder("context_str", "query_str"))
@@ -71,6 +94,6 @@ class SummarizeQueryConfig(BaseModel):
     content_format: ContentFormat = ContentFormat.PLAIN
 
 
-class SummarizeConfig(BaseModel):
-    chatgpt: ChatgptConfig
-    query: SummarizeQueryConfig
+class RagConfig(BaseModel):
+    retrieve: RetrieveConfig
+    summarize: SummarizeConfig

llm_backend/rag/content_formatters.py

@@ -0,0 +1,17 @@
+from collections.abc import Callable, Sequence
+from enum import StrEnum
+
+
+class ContentFormat(StrEnum):
+    PLAIN = "plain"
+    NUMBERED = "numbered"
+
+
+ContentFormatter = Callable[[Sequence[str]], Sequence[str]]
+
+CONTENT_FORMATTERS: dict[ContentFormat, ContentFormatter] = {
+    ContentFormat.PLAIN: lambda x: x,
+    ContentFormat.NUMBERED: lambda x: [
+        f"{i}. {line}" for i, line in enumerate(x, start=1)
+    ],
+}

llm_backend/rag/service.py

@@ -0,0 +1,25 @@
+import grpc
+
+from llm_backend.protos import rag_pb2, rag_pb2_grpc
+from llm_backend.rag.config import RagConfig
+from llm_backend.rag.workflow import RagWorkflow
+
+
+class RagService(rag_pb2_grpc.RagServiceServicer):
+    def __init__(self, config: RagConfig):
+        self.workflow = RagWorkflow(config=config)
+
+    async def Rag(
+        self,
+        request: rag_pb2.RagRequest,
+        context: grpc.aio.ServicerContext,
+    ):
+        result = await self.workflow.run(
+            keywords=request.keywords,
+            similarity_top_k=request.similarity_top_k,
+        )
+
+        return rag_pb2.RagResponse(
+            retrieved_ids=result["retrieved_ids"],
+            summary=result["summary"],
+        )