DST-199: base code for chunking

04-call-summaries/chunking.py

@@ -0,0 +1,134 @@
+import datetime
+from langchain_text_splitters import (
+ RecursiveCharacterTextSplitter,
+ NLTKTextSplitter,
+ SpacyTextSplitter,
+)
+from langchain_core.prompts import PromptTemplate
+from langchain.docstore.document import Document
+from llm import LLM
+from run import get_transcript
+
+
+# split text into chunks
+def get_text_chunks(text, chunk_size, chunk_overlap, text_splitter_choice):
+ if text_splitter_choice == "2":
+ text_splitter = NLTKTextSplitter()
+ elif text_splitter_choice == "3":
+ text_splitter = SpacyTextSplitter()
+ else:
+ text_splitter = RecursiveCharacterTextSplitter(
+ chunk_size=chunk_size, chunk_overlap=chunk_overlap
+ )
+
+ texts = text_splitter.split_text(text)
+
+ docs = [
+ Document(
+ page_content=t,
+ )
+ for t in texts
+ ]
+
+ return docs
+
+
+CHUNKING_PROMPT = """
+You are a helpful AI assistant tasked with summarizing transcripts, however we can only process the transcripts in pieces.
+Please fill out and return the following template: {template} with data in the text: {text}
+If the following template already has the field filled out, do not overwrite this information.
+"""
+
+initial_temp = """
+1. Caller Information:
+- Name
+- Contact Information
+- Availability
+- Household Information
+2. Reason/Type of Call: e.g., Applying for benefits, Follow-ups
+3. Previous Benefits History:
+- Applied for
+- Receives
+- Denied
+4. Benefits Discussion: Prefix the discussed benefit with a hashtag (e.g., #SNAP, #LIHEAP)
+5. Discussion Points:
+- Key information points
+6. Documents Needed: e.g., Income verification, Housing documentation
+7. Next Steps for Client
+8. Next Steps for Agent
+"""
+
+
+def chunking_ingest(transcript, prompt):
+ text_chunks = get_text_chunks(
+ transcript, chunk_size=750, chunk_overlap=300, text_splitter_choice="2"
+ )
+ prompt_template = PromptTemplate.from_template(prompt)
+ template = initial_temp
+
+ print("""
+ Select an llm
+ 1. openhermes (default)
+ 2. dolphin
+ 3. gemini
+ 4. gpt 4
+ 5. gpt 4o
+ 6. claude 3
+ """)
+
+ llm = input() or "1"
+
+ if llm == "2":
+ client = LLM(client_name="ollama", model_name="dolphin-mistral")
+ print("""----------
+ Dolphin
+ """)
+
+ elif llm == "3":
+ client = LLM(client_name="gemini")
+ print("""----------
+ Gemini Flash 1.5
+ """)
+ elif llm == "4":
+ print("""----------
+ GPT 4
+ """)
+ client = LLM(client_name="gpt", model_name="gpt4")
+ elif llm == "5":
+ print("""----------
+ GPT 4o
+ """)
+ client = LLM(client_name="gpt", model_name="gpt-4o")
+ elif llm == "6":
+ print("""----------
+ Claude 3
+ """)
+ client = LLM(client_name="claude")
+ else:
+ print("""
+ Openhermes
+ """)
+ client = LLM(client_name="ollama", model_name="openhermes")
+
+ client.init_client()
+ ct = datetime.datetime.now()
+ print("current time:-", ct)
+ for text in text_chunks:
+ formatted_prompt = prompt_template.format(
+ text=text.page_content, template=template
+ )
+ print("Processing Text Chunk")
+ template = client.generate_text(prompt=formatted_prompt)
+ print("Complete")
+ return template
+
+
+if __name__ == "__main__":
+ print(
+ chunking_ingest(
+ transcript=get_transcript("./son_calling_behalf_mother_transcript.txt"),
+ prompt=CHUNKING_PROMPT,
+ )
+ )
+ ct = datetime.datetime.now()
+ print(ct)

04-call-summaries/llm.py

@@ -9,79 +9,96 @@
 dotenv.load_dotenv()
 
 
-def get_transcript(file_path="./transcript.txt"):
- file = open(file_path, encoding="utf-8")
- content = file.read()
- return content
-
-
-def ollama_client(
- model_name=None,
- prompt=None,
- callbacks=None,
- settings=None,
-):
- if not settings:
- settings = {
- # "temperature": 0.1,
- # "system": "",
- # "template": "",
- # See https://github.com/langchain-ai/langchain/blob/master/libs/community/langchain_community/llms/ollama.py
- "stop": None
- }
-
- print("LLM settings:", model_name, settings)
- # To connect via another URL: Ollama(base_url='http://localhost:11434', ...)
- return Ollama(model=model_name, callbacks=callbacks, **settings).invoke(prompt)
-
-
-def google_gemini_client(
- model_name="gemini-pro",
- prompt=None,
- settings=None,
-):
- # Get a Google API key by following the steps after clicking on Get an API key button
- # at https://ai.google.dev/tutorials/setup
- GOOGLE_API_KEY = os.environ.get("GOOGLE_API_KEY")
-
- print("LLM settings:", model_name, settings)
-
- genai.configure(api_key=GOOGLE_API_KEY)
- if settings:
- genai.GenerationConfig(**settings)
- model = genai.GenerativeModel(model_name)
- return model.generate_content(prompt)
-
-
-def gpt3_5(prompt, model="gpt-3.5-turbo"):
- # Get API key from https://platform.openai.com/api-keys
- OPEN_AI_API_KEY = os.environ.get("OPEN_AI_API_KEY")
- openai_client = OpenAI(api_key=OPEN_AI_API_KEY) # Uses OPENAI_API_KEY
- return (
- openai_client.chat.completions.create(
- model=model, messages=[{"role": "user", "content": prompt}]
- )
- .choices[0]
- .message.content
- )
-
-
-def gpt_4_turbo(prompt):
- return gpt3_5(prompt, model="gpt-4-turbo")
-
-
-def claude(prompt, model="claude-3-opus-20240229", max_tokens=1024):
- # Get API key from https://console.anthropic.com/settings/keys
- ANTHROPIC_API_KEY = os.environ.get("ANTHROPIC_API_KEY")
-
- client = anthropic.Anthropic(
- api_key=ANTHROPIC_API_KEY,
- )
- generated_response = client.messages.create(
- model=model,
- max_tokens=max_tokens,
- messages=[{"role": "user", "content": prompt}],
- ).content
- text_response = "\n".join([text_block.text for text_block in generated_response])
-
- return text_response
+class LLM:
+ def __init__(
+ self,
+ client_name=None,
+ model_name=None,
+ max_tokens=1024,
+ settings=None,
+ ):
+ self.client_name = client_name
+ """Name of llm selection"""
+ self.model_name = model_name
+ """User friendly model name"""
+ self.model_version = model_name
+ """Exact model name being passed into the initializer"""
+ self.max_tokens = max_tokens
+ self.client = None
+ self.settings = settings
+
+ def init_client(self):
+ """Retrieves the llm client"""
+ if self.client_name == "ollama":
+ if self.settings is None:
+ self.settings = {
+ # "temperature": 0.1,
+ # "system": "",
+ # "template": "",
+ # See https://github.com/langchain-ai/langchain/blob/master/libs/community/langchain_community/llms/ollama.py
+ "stop": None
+ }
+ if self.model_name is None:
+ self.model_name = "openhermes"
+ # To connect via another URL: Ollama(base_url='http://localhost:11434', ...)
+ self.client = Ollama(model=self.model_version, **self.settings)
+
+ elif self.client_name == "gemini":
+ # Get a Google API key by following the steps after clicking on Get an API key button
+ # at https://ai.google.dev/tutorials/setup
+ GOOGLE_API_KEY = os.environ.get("GOOGLE_API_KEY")
+ if self.model_name == "gemini-pro":
+ self.model_version = "gemini-1.5-pro-latest"
+ else:
+ self.model_version = "gemini-1.5-flash-latest"
+
+ genai.configure(api_key=GOOGLE_API_KEY)
+ if self.settings is not None:
+ genai.GenerationConfig(**self.settings)
+ self.client = genai.GenerativeModel(self.model_version)
+
+ elif self.client_name == "gpt":
+ if self.model_name == "gpt3":
+ self.model_version = "gpt-3.5-turbo"
+ elif self.model_name == "gpt4":
+ self.model_version = "gpt-4-turbo"
+ else:
+ self.model_version = "gpt-4o"
+
+ # Get API key from https://platform.openai.com/api-keys
+ OPENAI_API_KEY = os.environ.get("OPENAI_API_KEY")
+ self.client = OpenAI(api_key=OPENAI_API_KEY) # Uses OPENAI_API_KEY
+
+ elif self.client_name == "claude":
+ self.model_version = "claude-3-opus-20240229"
+ ANTHROPIC_API_KEY = os.environ.get("ANTHROPIC_API_KEY")
+ self.client = anthropic.Anthropic(
+ api_key=ANTHROPIC_API_KEY,
+ )
+
+ def generate_text(self, prompt=None):
+ """Generates response given prompt"""
+ if self.client_name == "ollama":
+ return self.client.invoke(prompt)
+ elif self.client_name == "gemini":
+ return self.client.generate_content(prompt).text
+ elif self.client_name == "gpt":
+ return (
+ self.client.chat.completions.create(
+ model=self.model_version,
+ messages=[{"role": "user", "content": prompt}],
+ )
+ .choices[0]
+ .message.content
+ )
+ elif self.client_name == "claude":
+ generated_response = self.client.messages.create(
+ model=self.model_version,
+ max_tokens=self.max_tokens,
+ messages=[{"role": "user", "content": prompt}],
+ ).content
+ text_response = "\n".join(
+ [text_block.text for text_block in generated_response]
+ )
+
+ return text_response

04-call-summaries/requirements.txt

@@ -3,4 +3,9 @@ google-generativeai
 tokenizers
 langchain
 langchain_community
-openai
+openai
+sentence-transformers
+nltk
+spacy==3.7.4
+langchain-text-splitters
+langchain_openai