feat: update demo to pinsmart revamp

.env.example

@@ -1,4 +1,2 @@
-CHAT_RESPONSE_TOKEN="asd"
-CHAT_RESPONSE_URL="recombe-url"
-CACHE_RESPONSE_TOKEN="asd"
-CACHE_RESPONSE_URL="recom-calc-url"
+OPENAI_API_KEY=asdasdasd
+ES_URL=asdasdsad

app.py

@@ -1,106 +1,122 @@
-from http import HTTPStatus
+import json
+
+# Suppress all UserWarnings
+import warnings
+from typing import Any
 
 import gradio as gr
-import requests
+from langchain.retrievers import ContextualCompressionRetriever
+from langchain.retrievers.document_compressors import LLMChainFilter
+from langchain.schema import StrOutputParser
+from langchain_community.vectorstores.elasticsearch import ElasticsearchStore
+from langchain_core.documents import Document
+from langchain_core.prompts import ChatPromptTemplate
+from langchain_core.runnables import RunnablePassthrough
+from langchain_core.runnables.base import RunnableSerializable
+from langchain_openai import ChatOpenAI, OpenAIEmbeddings
 
 from config import get_settings
-from schemas import (
-    ChatbotQuestionAnswer,
-    PropertyKnowledgeIndexFAQRequest,
-    PropertyKnowledgeQuestionWithTokenObject,
-    PropertyKnowledgeAnswerObject,
+from prompts import CONTEXT_TEMPLATE, QA_WITH_CONTEXT_PROMPT, SYSTEM_PROMPT
+
+warnings.filterwarnings("ignore", category=UserWarning)
+
+KNOWLEDGE_BASE_INDEX_NAME = "pinsmart_revamp_prototype"
+EMBEDDING_MODEL = "text-embedding-3-large"
+RAG_LLM_MODEL = "gpt-3.5-turbo-0125"
+NROF_RETRIEVED_CONTEXT = 5
+
+ERROR_MESSAGE = "Terjadi kesalahan, silakan coba kembali"
+
+# ES only support up to 2048 embedding dimension
+embeddings = OpenAIEmbeddings(
+    model=EMBEDDING_MODEL,
+    api_key=get_settings().openai_api_key,
+    dimensions=2048,
 )
-import logging
-
-
-def index_faq(chatbot_qa_object: ChatbotQuestionAnswer):
-    data = PropertyKnowledgeIndexFAQRequest(
-        question=PropertyKnowledgeQuestionWithTokenObject(
-            token=chatbot_qa_object.question.token,
-            content=chatbot_qa_object.question.content,
-            embedding=chatbot_qa_object.question.embedding,
-            modelName=chatbot_qa_object.question.model_name,
-        ),
-        answer=PropertyKnowledgeAnswerObject(
-            content=chatbot_qa_object.answer.content,
-            categories=chatbot_qa_object.answer.categories,
-            source=chatbot_qa_object.answer.source,
-            referenceURLs=chatbot_qa_object.answer.reference_urls,
-            modelName=chatbot_qa_object.answer.model_name,
-        ),
-    )
 
-    faq_index_response = requests.post(
-        get_settings().cache_response_url,
-        headers={
-            "Authorization": "Bearer"
-            f" {get_settings().cache_response_token}",
-            "Content-Type": "application/json",
-        },
-        data=data.model_dump_json(by_alias=True),
-    )
-    if faq_index_response.status_code != HTTPStatus.OK:
-        logging.error(
-            f"failed to index FAQ, got error on: {faq_index_response.text}"
+vector_db = ElasticsearchStore(
+    es_url=get_settings().es_url,
+    index_name=KNOWLEDGE_BASE_INDEX_NAME,
+    embedding=embeddings,
+)
+
+llm = ChatOpenAI(
+    model_name=RAG_LLM_MODEL,
+    temperature=0,
+    api_key=get_settings().openai_api_key,
+)
+
+qa_prompt = ChatPromptTemplate.from_messages(
+    [
+        ("system", SYSTEM_PROMPT),
+        ("human", QA_WITH_CONTEXT_PROMPT),
+    ]
+)
+
+_filter = LLMChainFilter.from_llm(llm)
+docs_retriever = vector_db.as_retriever(
+    search_kwargs={"k": NROF_RETRIEVED_CONTEXT}
+)
+
+
+def format_contexts(docs: list[Document]) -> str:
+    aggregate_context = ""
+    for doc in docs:
+        context = CONTEXT_TEMPLATE.format(
+            source=doc.metadata["document_id"],
+            title=doc.metadata["heading_title"],
+            content=doc.page_content,
         )
 
+        aggregate_context += context
+
+    return aggregate_context
 
-def get_prompt_response(message: str):
-    data = {"question": message}
-    response = requests.post(
-        get_settings().chat_response_url,
-        headers={
-            "Authorization": "Bearer" f" {get_settings().chat_response_token}",
-            "Content-Type": "application/json",
-        },
-        json=data,
-    )
 
-    if response.status_code == HTTPStatus.OK:
-        try:
-            chatbot_qa_object = ChatbotQuestionAnswer.model_validate(
-                response.json()["data"]
-            )
-        except ValueError as val_err:
-            logging.error(
-                f"got value on chatbot response: {response.text},"
-                f"error message: {val_err}"
-            )
+docs_compression_retriever = ContextualCompressionRetriever(
+    base_compressor=_filter, base_retriever=docs_retriever
+)
 
-            return "Terjadi kesalahan, silakan coba kembali"
 
-        except Exception as e:
-            logging.error(
-                f"got unknown error on chabot find answer request, error: {e}"
-            )
+def get_end_device_target(*args: Any) -> str:
+    return "mobile"
 
-            return "Terjadi kesalahan, silakan coba kembali"
 
-        if chatbot_qa_object.question.embedding is not None:
-            try:
-                index_faq(chatbot_qa_object)
-            except Exception as e:
-                logging.error(f"failed to index FAQ, error: {e}")
+rag_chain: RunnableSerializable = (
+    {
+        "context": docs_compression_retriever | format_contexts,
+        "device": get_end_device_target,
+        "question": RunnablePassthrough(),
+    }
+    | qa_prompt
+    | llm
+    | StrOutputParser()
+)
 
-        if len(chatbot_qa_object.answer.reference_urls) > 0:
-            bot_msg = chatbot_qa_object.answer.content + "\n\n" + "Sumber: \n"
 
-            for source in chatbot_qa_object.answer.reference_urls:
-                bot_msg = bot_msg + f"- {source}\n"
+def parse_response(response: str) -> tuple[str, list[dict]]:
+    answer = response.split("[ANSWER]")[1].split("[RELATED DOCS]")[0].strip()
 
-            return bot_msg
+    related_docs = json.loads(response.split("[RELATED DOCS]")[1].strip())
 
-        return chatbot_qa_object.answer.content
+    return answer, related_docs
 
-    if response.status_code == HTTPStatus.NOT_FOUND:
-        return "Maaf, saya tidak dapat menemukan jawaban untuk pertanyaan ini."
 
-    logging.error(
-        f"got status code: {response.status_code}"
-        f"and error message: {response.text}"
-    )
+def get_prompt_response(message: str):
+    response = rag_chain.invoke(message)
+
+    try:
+        answer, related_docs = parse_response(response)
+    except Exception as e:
+        print(f"Error : {e}")
+        return ERROR_MESSAGE
+
+    reference_str = "Referensi: \n"
+    for doc in related_docs:
+        reference_title = doc["title"]
+        reference_str += f"- Artikel: {reference_title}\n"
 
-    return "Terjadi kesalahan, silakan coba kembali"
+    return answer + "\n\n" + reference_str
 
 
 def chat(message, history):
@@ -124,7 +140,7 @@ with gr.Blocks() as demo:
         "
         >
         <h1 style="font-weight: 900; margin-bottom: 7px; margin-top: 5px;">
-            PinGPT
+            PinSmart Prototype
         </h1>
         </div>
     </div>"""
@@ -133,7 +149,7 @@ with gr.Blocks() as demo:
     with gr.Row():
         question = gr.Textbox(
             label=(
-                "Masukkan pertanyaan seputar property (KPR, Take Over KPR,"
+                "Masukkan pertanyaan seputar properti (KPR, Take Over KPR,"
                 " dsb)",
             ),
             placeholder="Apa dokumen yang diperlukan untuk KPR?",

config.py

@@ -4,13 +4,11 @@ from pydantic_settings import BaseSettings, SettingsConfigDict
 
 
 class Settings(BaseSettings):
-    chat_response_url: str
-    chat_response_token: str
-    cache_response_url: str
-    cache_response_token: str
+    openai_api_key: str
+    es_url: str
 
     model_config = SettingsConfigDict(
-        env_file=".env", env_file_encoding="utf-8"
+        env_file=".env", env_file_encoding="utf-8", extra="ignore"
     )
 
 

prompts.py

@@ -0,0 +1,68 @@
+SYSTEM_PROMPT = (
+    "You are a world class property consultant for Indonesia market recognized"
+    " as PinSmart. You work for Pinhome.id and hence cannot mention other "
+    "competitor such as rumah.com, rumah123.com, lamudi.id, 99.co, and so on."
+)
+QA_WITH_CONTEXT_PROMPT = """
+## Instructions
+
+- Always generate answer only based on the contexts and facts
+- Take your time to generate the answer and elaborate all the supporting
+    details in the answer
+- The generated response MUST use the following format:
+    ```
+    [ANSWER]
+    put the answer here
+
+    [RELATED DOCS]
+    [
+        {{
+            "source" : "context_source1",
+            "title" : "context_title1"
+        }},
+        {{
+            "source" : "context_source2",
+            "title" : "context_title2"
+        }},
+        ...
+    ]
+    ```
+- Only show [RELATED DOCS] that is exist on the provided context and do
+    not show duplicate [RELATED DOCS]
+- Always generate [ANSWER] in Indonesian language
+- Generate the [ANSWER] using the following formatting rules:
+    - Must be easy to read and using good structure such as
+        bullet and numbering if necessary
+    - Use the best string formatting if the [ANSWER] is shown in
+        {device} app as chatbot app
+
+## Contexts
+{context}
+
+## Question
+{question}
+
+## Answer
+"""
+
+CONTEXT_TEMPLATE = """
+---
+### Context Metadata
+source: {source}
+title: {title}
+
+### Content
+
+{content}
+"""
+
+CONTEXT_TEMPLATE = """
+---
+### Context Metadata
+source: {source}
+title: {title}
+
+### Content
+
+{content}
+"""

pyproject.toml

@@ -10,6 +10,10 @@ packages = []
 python = "^3.10"
 gradio = { version = "3.44.3", source = "pypi" }
 pydantic-settings = { version = "^2.0.3", extras = ["dotenv"], source = "pypi" }
+langchain = {version = "^0.1.7", source = "pypi"}
+langchain-openai = {version = "^0.0.6", source = "pypi"}
+tiktoken = {version = "^0.6.0", source = "pypi"}
+elasticsearch = {version = "^8.12.0", source = "pypi"}
 
 
 [tool.ruff]

requirements.txt

@@ -1,7 +1,10 @@
 aiofiles==23.2.1 ; python_version >= "3.10" and python_version < "4.0"
+aiohttp==3.9.3 ; python_version >= "3.10" and python_version < "4.0"
+aiosignal==1.3.1 ; python_version >= "3.10" and python_version < "4.0"
 altair==5.1.1 ; python_version >= "3.10" and python_version < "4.0"
 annotated-types==0.5.0 ; python_version >= "3.10" and python_version < "4.0"
 anyio==4.0.0 ; python_version >= "3.10" and python_version < "4.0"
+async-timeout==4.0.3 ; python_version >= "3.10" and python_version < "3.11"
 attrs==23.1.0 ; python_version >= "3.10" and python_version < "4.0"
 certifi==2023.7.22 ; python_version >= "3.10" and python_version < "4.0"
 charset-normalizer==3.2.0 ; python_version >= "3.10" and python_version < "4.0"
@@ -9,14 +12,20 @@ click==8.1.7 ; python_version >= "3.10" and python_version < "4.0"
 colorama==0.4.6 ; python_version >= "3.10" and python_version < "4.0" and platform_system == "Windows"
 contourpy==1.1.1 ; python_version >= "3.10" and python_version < "4.0"
 cycler==0.11.0 ; python_version >= "3.10" and python_version < "4.0"
+dataclasses-json==0.6.4 ; python_version >= "3.10" and python_version < "4.0"
+distro==1.9.0 ; python_version >= "3.10" and python_version < "4.0"
+elastic-transport==8.12.0 ; python_version >= "3.10" and python_version < "4.0"
+elasticsearch==8.12.0 ; python_version >= "3.10" and python_version < "4.0"
 exceptiongroup==1.1.3 ; python_version >= "3.10" and python_version < "3.11"
 fastapi==0.103.0 ; python_version >= "3.10" and python_version < "4.0"
 ffmpy==0.3.1 ; python_version >= "3.10" and python_version < "4.0"
 filelock==3.12.4 ; python_version >= "3.10" and python_version < "4.0"
 fonttools==4.42.1 ; python_version >= "3.10" and python_version < "4.0"
+frozenlist==1.4.1 ; python_version >= "3.10" and python_version < "4.0"
 fsspec==2023.9.1 ; python_version >= "3.10" and python_version < "4.0"
 gradio-client==0.5.0 ; python_version >= "3.10" and python_version < "4.0"
 gradio==3.44.3 ; python_version >= "3.10" and python_version < "4.0"
+greenlet==3.0.3 ; python_version >= "3.10" and python_version < "4.0" and (platform_machine == "win32" or platform_machine == "WIN32" or platform_machine == "AMD64" or platform_machine == "amd64" or platform_machine == "x86_64" or platform_machine == "ppc64le" or platform_machine == "aarch64")
 h11==0.14.0 ; python_version >= "3.10" and python_version < "4.0"
 httpcore==0.18.0 ; python_version >= "3.10" and python_version < "4.0"
 httpx==0.25.0 ; python_version >= "3.10" and python_version < "4.0"
@@ -24,14 +33,25 @@ huggingface-hub==0.17.1 ; python_version >= "3.10" and python_version < "4.0"
 idna==3.4 ; python_version >= "3.10" and python_version < "4.0"
 importlib-resources==6.0.1 ; python_version >= "3.10" and python_version < "4.0"
 jinja2==3.1.2 ; python_version >= "3.10" and python_version < "4.0"
+jsonpatch==1.33 ; python_version >= "3.10" and python_version < "4.0"
+jsonpointer==2.4 ; python_version >= "3.10" and python_version < "4.0"
 jsonschema-specifications==2023.7.1 ; python_version >= "3.10" and python_version < "4.0"
 jsonschema==4.19.0 ; python_version >= "3.10" and python_version < "4.0"
 kiwisolver==1.4.5 ; python_version >= "3.10" and python_version < "4.0"
+langchain-community==0.0.20 ; python_version >= "3.10" and python_version < "4.0"
+langchain-core==0.1.23 ; python_version >= "3.10" and python_version < "4.0"
+langchain-openai==0.0.6 ; python_version >= "3.10" and python_version < "4.0"
+langchain==0.1.7 ; python_version >= "3.10" and python_version < "4.0"
+langsmith==0.0.87 ; python_version >= "3.10" and python_version < "4.0"
 markupsafe==2.1.3 ; python_version >= "3.10" and python_version < "4.0"
+marshmallow==3.20.2 ; python_version >= "3.10" and python_version < "4.0"
 matplotlib==3.8.0 ; python_version >= "3.10" and python_version < "4.0"
+multidict==6.0.5 ; python_version >= "3.10" and python_version < "4.0"
+mypy-extensions==1.0.0 ; python_version >= "3.10" and python_version < "4.0"
 numpy==1.25.2 ; python_version >= "3.10" and python_version < "4.0"
+openai==1.12.0 ; python_version >= "3.10" and python_version < "4.0"
 orjson==3.9.7 ; python_version >= "3.10" and python_version < "4.0"
-packaging==23.1 ; python_version >= "3.10" and python_version < "4.0"
+packaging==23.2 ; python_version >= "3.10" and python_version < "4.0"
 pandas==2.1.0 ; python_version >= "3.10" and python_version < "4.0"
 pillow==10.0.1 ; python_version >= "3.10" and python_version < "4.0"
 pydantic-core==2.6.3 ; python_version >= "3.10" and python_version < "4.0"
@@ -45,6 +65,7 @@ python-multipart==0.0.6 ; python_version >= "3.10" and python_version < "4.0"
 pytz==2023.3.post1 ; python_version >= "3.10" and python_version < "4.0"
 pyyaml==6.0.1 ; python_version >= "3.10" and python_version < "4.0"
 referencing==0.30.2 ; python_version >= "3.10" and python_version < "4.0"
+regex==2023.12.25 ; python_version >= "3.10" and python_version < "4.0"
 requests==2.31.0 ; python_version >= "3.10" and python_version < "4.0"
 rpds-py==0.10.3 ; python_version >= "3.10" and python_version < "4.0"
 semantic-version==2.10.0 ; python_version >= "3.10" and python_version < "4.0"
@@ -52,12 +73,17 @@ setuptools-scm==7.1.0 ; python_version >= "3.10" and python_version < "4.0"
 setuptools==68.2.2 ; python_version >= "3.10" and python_version < "4.0"
 six==1.16.0 ; python_version >= "3.10" and python_version < "4.0"
 sniffio==1.3.0 ; python_version >= "3.10" and python_version < "4.0"
+sqlalchemy==2.0.27 ; python_version >= "3.10" and python_version < "4.0"
 starlette==0.27.0 ; python_version >= "3.10" and python_version < "4.0"
+tenacity==8.2.3 ; python_version >= "3.10" and python_version < "4.0"
+tiktoken==0.6.0 ; python_version >= "3.10" and python_version < "4.0"
 tomli==2.0.1 ; python_version >= "3.10" and python_version < "3.11"
 toolz==0.12.0 ; python_version >= "3.10" and python_version < "4.0"
 tqdm==4.66.1 ; python_version >= "3.10" and python_version < "4.0"
 typing-extensions==4.7.1 ; python_version >= "3.10" and python_version < "4.0"
+typing-inspect==0.9.0 ; python_version >= "3.10" and python_version < "4.0"
 tzdata==2023.3 ; python_version >= "3.10" and python_version < "4.0"
 urllib3==2.0.4 ; python_version >= "3.10" and python_version < "4.0"
 uvicorn==0.23.2 ; python_version >= "3.10" and python_version < "4.0"
 websockets==11.0.3 ; python_version >= "3.10" and python_version < "4.0"
+yarl==1.9.4 ; python_version >= "3.10" and python_version < "4.0"

schemas.py

@@ -1,77 +0,0 @@
-from typing import Optional
-from pydantic import BaseModel, Field, validator
-from enum import Enum
-
-
-class ChatbotQuestion(BaseModel):
-    token: str = Field(..., alias="token")
-    content: str = Field(..., alias="content")
-    embedding: Optional[list[float]] = Field(None, alias="embedding")
-    model_name: str = Field(..., alias="modelName")
-
-
-class ChatbotAnswer(BaseModel):
-    token: str = Field(..., alias="token")
-    content: str = Field(..., alias="content")
-    reference_urls: list[str] = Field(..., alias="referenceURLs")
-    categories: list[str] = Field(..., alias="categories")
-    source: str = Field(..., alias="source")
-    model_name: str = Field(..., alias="modelName")
-
-
-class ChatbotQuestionAnswer(BaseModel):
-    question: ChatbotQuestion = Field(..., alias="question")
-    answer: ChatbotAnswer = Field(..., alias="answer")
-
-
-class CategoriesObject(Enum):
-    property_seekers = "property_seekers"
-    property_owners = "property_owners"
-    mortgage = "mortgage"
-
-
-class SourcesObject(Enum):
-    manual = "manual"
-    chatbot = "chatbot"
-    knowledge_docs_unavailable = "knowledge_docs_unavailable"
-
-
-class PropertyKnowledgeAnswerObject(BaseModel):
-    content: str
-    categories: list[CategoriesObject] = Field([])
-    source: SourcesObject
-    reference_urls: list[str] = Field([], alias="referenceURLs")
-    model_name: Optional[str] = Field(None, alias="modelName")
-
-    @validator("content")
-    def validate_answer_content(cls, content: str) -> dict:
-        if content == "":
-            raise ValueError("content can't be empty string")
-
-        return content
-
-    @validator("model_name")
-    def validate_answer_model_name(cls, model_name: str) -> dict:
-        if model_name is not None and model_name == "":
-            return None
-
-        return model_name
-
-
-class PropertyKnowledgeQuestionWithTokenObject(BaseModel):
-    token: Optional[str]
-    content: str
-    embedding: Optional[list[float]] = Field([])
-    model_name: Optional[str] = Field(None, alias="modelName")
-
-    @validator("content")
-    def validate_question_content(cls, content: str) -> dict:
-        if content == "":
-            raise ValueError("content can't be empty string")
-
-        return content
-
-
-class PropertyKnowledgeIndexFAQRequest(BaseModel):
-    question: PropertyKnowledgeQuestionWithTokenObject
-    answer: PropertyKnowledgeAnswerObject