feature(#8): LLM Strategy 1 - Build our falcon model

src/common/utils.py

@@ -23,7 +23,7 @@ PINECONE_NAMESPACE = "risinglangchain-namespace"
 PINECONE_INDEX_NAME = "risinglangchain-index"
 
 # open ai
-GPT_MODEL = "gpt-3.5-turbo"
+DEFAULT_GPT_MODEL = "gpt-4"
 
 # AI Agent name
 AGENT_NAME = "RisingBrain Assistant"

src/rising_plugin/guardrails-config/actions/actions.py

@@ -46,6 +46,7 @@ from src.rising_plugin.llm.llms import (
 
 @action()
 async def general_question(query, model, uuid, image_search):
+    """step1: handle with gpt-4"""
     file_path = os.path.dirname(os.path.abspath(__file__))
 
     with open(f"{file_path}/phone.json", "r") as infile:
@@ -68,13 +69,13 @@ async def general_question(query, model, uuid, image_search):
 
     chain_data = get_llm_chain(model=model).run(input_documents=docs, question=query)
     # test
-    if model == GPT_3_5_TURBO or model == GPT_4 or model == GPT_4_32K:
-        gpt_llm = GptLLM(model=model)
-        chain_data = gpt_llm.get_chain().run(input_documents=docs, question=query)
-    elif model == FALCON_7B:
-        falcon_llm = FalconLLM()
-        chain_data = falcon_llm.get_chain().run(input_documents=docs, question=query)
-
+    # if model == GPT_3_5_TURBO or model == GPT_4 or model == GPT_4_32K:
+    #     gpt_llm = GptLLM(model=model)
+    #     chain_data = gpt_llm.get_chain().run(input_documents=docs, question=query)
+    # elif model == FALCON_7B:
+    #     falcon_llm = FalconLLM()
+    #     chain_data = falcon_llm.get_chain().run(question=query)
+    falcon_llm = FalconLLM()
     try:
         result = json.loads(chain_data)
         # check image query with only its text
@@ -86,6 +87,9 @@ async def general_question(query, model, uuid, image_search):
 
             # else:
             #     return result
+        """check program is message to handle it with falcon llm"""
+        if result["program"] == "message":
+            result["content"] = falcon_llm.query(question=query)
         return str(result)
     except ValueError as e:
         # Check sms and browser query
@@ -93,4 +97,4 @@ async def general_question(query, model, uuid, image_search):
             return str({"program": "sms", "content": chain_data})
         elif doc_list[0] in COMMAND_BROWSER_OPEN:
             return str({"program": "browser", "content": "https://google.com"})
-        return str({"program": "message", "content": chain_data})
+        return str({"program": "message", "content": falcon_llm.query(question=query)})

src/rising_plugin/guardrails-config/actions/phone.csv

@@ -6,15 +6,15 @@
 "If user said that send sms or text, please answer belowing json format. {""program"": ""sms"", ""content"": ""ask who""}"
 "If user said that compose, write, or create an sms message, please answer belowing json format. {""program"": ""sms"", ""content"": ""ask who""}"
 "If user said that search contact with its description such as display name or phone number, please answer belowing json format. {""program"": ""contact"", ""content"": ""description of the contact that user is going to search""}"
-"If user said that open a new tab, go to a new tab, or open a new page in a web browser, please answer belowing json format. {""program"": ""open_tab"", ""content"": """"}"
-"If user said that open a new tab and search, go to a new tab and search in a web browser, or open a new page and search, please answer belowing json format. {""program"": ""open_tab_search"", ""content"": ""keyword that user is going to search""}"
-"If user said that close a tab, please answer belowing json format in a web browser. {""program"": ""close_tab"", ""content"": """"}"
-"If user said that launch a browser or open a browser in a web browser, please answer belowing json format. {""program"": ""browser"", ""content"": ""https://google.com""}"
-"If user said that go to a previous page, or open a previous page in a web browser, please answer belowing json format. {""program"": ""previous_page"", ""content"": """"}"
-"If user said that go to a next page, or open a next page in a web browser, please answer belowing json format. {""program"": ""next_page"", ""content"": """"}"
-"If user said that scroll up, scroll up page, or page scroll up in a web browser, please answer belowing json format. {""program"": ""scroll_up"", ""content"": """"}"
-"If user said that scroll down, scroll down page, page scroll down in a web browser, please answer belowing json format. {""program"": ""scroll_down"", ""content"": """"}"
-"If user said that scroll top, scroll top page, or scroll top of page in a web browser, please answer belowing json format. {""program"": ""scroll_top"", ""content"": """"}"
-"If user said that scroll bottom, scroll bottom page, or scroll bottom of page in a web browser, please answer belowing json format. {""program"": ""scroll_bottom"", ""content"": """"}"
+"If user said that open a tab, go to a tab, or open a page, please answer belowing json format. {""program"": ""open_tab"", ""content"": """"}"
+"If user said that open a tab and search, go to a tab and search, or open a page and search, please answer belowing json format. {""program"": ""open_tab_search"", ""content"": ""keyword that user is going to search""}"
+"If user said that close a tab, please answer belowing json format. {""program"": ""close_tab"", ""content"": """"}"
+"If user said that launch a browser or open a browser, please answer belowing json format. {""program"": ""browser"", ""content"": ""https://google.com""}"
+"If user said that go to a previous page, or open a previous page, please answer belowing json format. {""program"": ""previous_page"", ""content"": """"}"
+"If user said that go to a next page, or open a next page, please answer belowing json format. {""program"": ""next_page"", ""content"": """"}"
+"If user said that scroll up, scroll up page, or page scroll up, please answer belowing json format. {""program"": ""scroll_up"", ""content"": """"}"
+"If user said that scroll down, scroll down page, page scroll down, please answer belowing json format. {""program"": ""scroll_down"", ""content"": """"}"
+"If user said that scroll top, scroll top page, or scroll top of page, please answer belowing json format. {""program"": ""scroll_top"", ""content"": """"}"
+"If user said that scroll bottom, scroll bottom page, or scroll bottom of page, please answer belowing json format. {""program"": ""scroll_bottom"", ""content"": """"}"
 "If user is going to select an item, an article or a website with its description in a web browser, please answer belowing json format. {""program"": ""select_item_detail_info"", ""content"": ""the description of an item, an article or a website in a browser""}"
 "If all of above is not correct, please give the most appropriate answer to the user's question. Please answer belowing json format. {""program"":""message"", ""content"":""your answer""}"

src/rising_plugin/guardrails-config/general.co

@@ -15,5 +15,5 @@ define bot inform capabilities
 define flow
   priority 0.9
   user ...
-  $result = execute general_question(query=$last_user_message, model="gpt-3.5-turbo", uuid="", image_search=True)
+  $result = execute general_question(query=$last_user_message, model="gpt-4", uuid="", image_search=True)
   bot $result

src/rising_plugin/llm/falcon_llm.py

@@ -27,3 +27,9 @@ class FalconLLM:
         prompt = PromptTemplate(template=template, input_variables=["question"])
         llm_chain = LLMChain(prompt=prompt, llm=self.llm, verbose=True)
         return llm_chain
+
+    """getting the output in query with falcon llm"""
+
+    def query(self, question: str) -> str:
+        chain = self.get_chain()
+        return chain.run(question=question)

src/rising_plugin/risingplugin.py

@@ -7,7 +7,6 @@ import textwrap
 
 from typing import Any
 
-from langchain.chains.question_answering import load_qa_chain
 from nemoguardrails.rails import LLMRails, RailsConfig
 
 from langchain.chat_models import ChatOpenAI
@@ -18,7 +17,6 @@ from .llm.llms import get_llm, GPT_4, FALCON_7B
 from ..common.utils import (
     OPENAI_API_KEY,
     FIREBASE_STORAGE_ROOT,
-    parseJsonFromCompletion,
 )
 from .image_embedding import (
     query_image_text,
@@ -50,8 +48,24 @@ def processLargeText(app: any, chunks: any):
             ]
         )
         result = json.dumps(message["content"])
-
-        return parseJsonFromCompletion(result)
+        result = result[1:-1]
+        # fmt: off
+        result = result.replace("{'", '{"')
+        result = result.replace("'}", '"}')
+        result = result.replace("': '", '": "')
+        result = result.replace("': \\\"", '": \"')
+        result = result.replace("', '", '", "')
+
+        substring = '\\"}'
+        replacement = '\"}'
+
+        index = result.rfind(substring)
+
+        if index == len(result) - 3:
+            result = result[:index] + replacement + result[index + len(substring):]
+        # fmt: on
+        result = json.loads(result)
+        return result
     else:
         first_query = "The total length of the content that I want to send you is too large to send in only one piece.\nFor sending you that content, I will follow this rule:\n[START PART 1/10]\nThis is the content of the part 1 out of 10 in total\n[END PART 1/10]\nThen you just answer: 'Received part 1/10'\nAnd when I tell you 'ALL PART SENT', then you can continue processing the data and answering my requests."
         app.generate(messages=[{"role": "user", "content": first_query}])
@@ -103,7 +117,24 @@ def processLargeText(app: any, chunks: any):
                     messages=[{"role": "user", "content": last_query}]
                 )
                 result = json.dumps(message["content"])
-                return parseJsonFromCompletion(result)
+                result = result[1:-1]
+                # fmt: off
+                result = result.replace("{'", '{"')
+                result = result.replace("'}", '"}')
+                result = result.replace("': '", '": "')
+                result = result.replace("': \\\"", '": \"')
+                result = result.replace("', '", '", "')
+
+                substring = '\\"}'
+                replacement = '\"}'
+
+                index = result.rfind(substring)
+
+                if index == len(result) - 3:
+                    result = result[:index] + replacement + result[index + len(substring):]
+                # fmt: on
+                result = json.loads(result)
+                return result
         # out of for-loop
 
 
@@ -122,17 +153,6 @@ def getCompletion(
     return processLargeText(app, chunks)
 
 
-def getCompletionOnly(
-    query: str,
-    model: str = "gpt-4",
-) -> str:
-    llm = ChatOpenAI(model_name=model, temperature=1.7, openai_api_key=OPENAI_API_KEY)
-    chain = load_qa_chain(llm, chain_type="stuff")
-    test_question = """Please return the link of best relatedness of item which the title is "Android Studio in browser" from the below data. [{"title": "Android Studio", "link": "https://android.com"} , {"title": "What's this?", "link": "https://test.com"} , {"title": "How are you?", "link": "https://d.com"}]"""
-    chain_data = chain.run(input_documents=[], question=test_question)
-    return chain_data
-
-
 def query_image_ask(image_content, message, uuid):
     prompt_template = get_prompt_image_with_message(image_content, message)
     try:

src/service/llm/chat_service.py

@@ -3,7 +3,7 @@ import time
 
 from openai.error import RateLimitError
 
-from src.common.utils import AGENT_NAME, GPT_MODEL
+from src.common.utils import AGENT_NAME, DEFAULT_GPT_MODEL
 from src.rising_plugin.risingplugin import handle_chat_completion
 from src.logs import logger
 from src.model.chat_response_model import ChatResponseModel
@@ -11,7 +11,7 @@ from src.model.message_model import MessageModel
 
 
 class ChatService:
-    def __init__(self, ai_name=AGENT_NAME, llm_model=GPT_MODEL):
+    def __init__(self, ai_name=AGENT_NAME, llm_model=DEFAULT_GPT_MODEL):
         self.ai_name = ai_name
         self.llm_model = llm_model