Merge pull request #358 from artificialfintelligence/blogger

Blog-Writer Feature

demo/blog_writer/.env.sample

@@ -0,0 +1,6 @@
+# Enter your OpenAI API key and Serper API key immediately after the equals sign without spaces or enclosing quotation marks.
+OPENAI_API_KEY=
+SERPER_API_KEY=
+
+# Uncomment the line below to enable debugging logs.
+# LOG_LEVEL=DEBUG

demo/blog_writer/README.md

@@ -0,0 +1,5 @@
+# Blog Writer Project
+This community project aims to utilize Sherpa as a library (`sherpa_ai`) as well as direct LLM calls (to the OpenAI API) with LangChain to construct a blog post from the raw 
+transcript of a lecture or presenttaion.  
+
+Please refer to [this How-to Guide](https://github.com/Aggregate-Intellect/sherpa/tree/main/docs/How_To/Tutorials/blog_writer.rst) for further details.

demo/blog_writer/actions.py

@@ -0,0 +1,55 @@
+from langchain.document_loaders import PDFMinerLoader
+from langchain.text_splitter import SentenceTransformersTokenTextSplitter
+from langchain.vectorstores.chroma import Chroma
+from loguru import logger
+
+from sherpa_ai.actions.base import BaseAction
+
+
+class DocumentSearch(BaseAction):
+    def __init__(self, filename, embedding_function, k=4):
+        # file name of the pdf
+        self.filename = filename
+        # the embedding function to use
+        self.embedding_function = embedding_function
+        # number of results to return in search
+        self.k = k
+
+        # load the pdf and create the vector store
+        self.chroma = Chroma(embedding_function = embedding_function)
+        documents = PDFMinerLoader(self.filename).load()
+        documents = SentenceTransformersTokenTextSplitter(chunk_overlap=0).split_documents(documents)
+
+        logger.info(f"Adding {len(documents)} documents to the vector store")
+        self.chroma.add_documents(documents)
+        logger.info("Finished adding documents to the vector store")
+
+    def execute(self, query):
+        """
+        Execute the action by searching the document store for the query
+
+        Args:
+            query (str): The query to search for
+
+        Returns:
+            str: The search results combined into a single string
+        """
+
+        results = self.chroma.search(query, search_type="mmr", k=self.k)
+        return "\n\n".join([result.page_content for result in results])
+
+    @property
+    def name(self) -> str:
+        """
+        The name of the action, used to describe the action to the agent.
+        """
+        return "DocumentSearch"
+
+    @property
+    def args(self) -> dict:
+        """
+        The arguments that the action takes, used to describe the action to the agent.
+        """
+        return {
+            "query": "string"
+        }

demo/blog_writer/agent_config.yml

@@ -0,0 +1,51 @@
+shared_memory:
+    _target_: sherpa_ai.memory.shared_memory.SharedMemory  # The absolute path to the share memory class in the library
+    objective: Answer the question  # Objective for the agent, since this is a question answering agent, the objective is to answer questions
+
+agent_config: # For the demo, default configuration is used. You can change the configuration as per your requirement
+    _target_: sherpa_ai.config.task_config.AgentConfig
+
+
+llm:  # Configuration for the llm, here we are using the OpenAI GPT-3.5-turbo model
+    _target_: langchain.chat_models.ChatOpenAI
+    model_name: gpt-3.5-turbo
+    temperature: 0
+
+embedding_func:
+    _target_: langchain.embeddings.SentenceTransformerEmbeddings
+    model_name: sentence-transformers/all-mpnet-base-v2
+
+doc_search:
+    _target_: actions.DocumentSearch
+    filename: transcript.pdf
+    embedding_function: ${embedding_func}
+    k: 4
+
+google_search:
+    _target_: sherpa_ai.actions.GoogleSearch
+    role_description: Act as a question answering agent
+    task: Question answering
+    llm: ${llm}
+    include_metadata: true
+    config: ${agent_config}
+
+citation_validation:  # The tool used to validate and add citation to the answer
+    _target_: sherpa_ai.output_parsers.citation_validation.CitationValidation
+    sequence_threshold: 0.6
+    jaccard_threshold: 0.6
+    token_overlap: 0.6
+
+qa_agent:
+    _target_: sherpa_ai.agents.qa_agent.QAAgent
+    llm: ${llm}
+    shared_memory: ${shared_memory}
+    name: QA Sherpa
+    description: You are a technical writing assistant that helps users write articles. For each prompt, use Google Search to find detailed information that supports and expands on the prompt.
+    agent_config: ${agent_config}
+    num_runs: 1
+    validation_steps: 1
+    actions:
+        - ${google_search}
+        # - ${doc_search}
+    validations:
+        - ${citation_validation}

demo/blog_writer/main.py

@@ -0,0 +1,87 @@
+import json
+from argparse import ArgumentParser
+
+from hydra.utils import instantiate
+from omegaconf import OmegaConf
+from sherpa_ai.agents import QAAgent
+from sherpa_ai.events import EventType
+
+from outliner import Outliner
+
+# from sherpa_ai.memory import Belief
+
+
+def get_qa_agent_from_config_file(
+    config_path: str,
+) -> QAAgent:
+    """
+    Create a QAAgent from a config file.
+
+    Args:
+        config_path: Path to the config file
+
+    Returns:
+        QAAgent: A QAAgent instance
+    """
+
+    config = OmegaConf.load(config_path)
+
+    agent_config = instantiate(config.agent_config)
+    qa_agent: QAAgent = instantiate(config.qa_agent, agent_config=agent_config)
+
+    return qa_agent
+
+
+if __name__ == "__main__":
+    parser = ArgumentParser()
+    parser.add_argument("--config", type=str, default="agent_config.yaml")
+    parser.add_argument("--transcript", type=str, default="transcript.txt")
+    args = parser.parse_args()
+
+    writer_agent = get_qa_agent_from_config_file(args.config)
+
+    outliner = Outliner(args.transcript)
+    blueprint = outliner.full_transcript2outline_json(verbose=True)
+    if blueprint.startswith("```"):
+        # The first and last lines are code block delimiters; remove them
+        lines = blueprint.split("\n")[1:-1]
+        pure_json_str = "\n".join(lines)
+    else:
+        pure_json_str = blueprint
+
+    with open("blueprint.json", "w") as f:
+        f.write(pure_json_str)
+
+    # with open("blueprint_manual.json", "r") as f:
+    #     pure_json_str = f.read()
+
+    parsed_json = json.loads(pure_json_str)
+
+    blog = ""
+    thesis = parsed_json.get("Thesis Statement", "")
+    blog += f"# Introduction\n{thesis}\n"
+    arguments = parsed_json.get("Supporting Arguments", [])
+    for argument in arguments:
+        blog += f"## {argument['Argument']}\n"
+        evidences = argument.get("Evidence", [])
+        for evidence in evidences:
+            writer_agent.shared_memory.add(EventType.task, "human", evidence)
+            result = writer_agent.run()
+            # writer_agent.belief = Belief()
+            blog += f"{result}\n"
+
+    with open("blog.md", "w") as f:
+        f.write(blog)
+
+    print("\nBlog generated successfully!\n")
+
+    # save_format = None
+    # while save_format is None:
+    #     save_format = input(
+    #         "Select format to save the blog in: 1. Markdown (Default) 2. ReStructured Text\n"
+    #     )
+
+    # if save_format == "2":
+    #     output = pypandoc.convert("blog.md", "rst")
+    #     if os.path.exists("blog.md"):
+    #         os.remove("blog.md")

demo/blog_writer/outliner.py

@@ -0,0 +1,148 @@
+import os
+import time
+
+import tiktoken
+from langchain.chat_models import ChatOpenAI
+from langchain.prompts.chat import (
+    ChatPromptTemplate,
+    HumanMessagePromptTemplate,
+    SystemMessagePromptTemplate,
+)
+from langchain.text_splitter import MarkdownTextSplitter
+
+
+class Outliner:
+    def __init__(self, transcript_file) -> None:
+        with open(transcript_file, "r") as f:
+            self.raw_transcript = f.read()
+        # instantiate chat model
+        self.chat = ChatOpenAI(
+            openai_api_key=os.environ.get("OPENAI_API_KEY"),
+            temperature=0,
+            model="gpt-3.5-turbo",
+        )
+
+    def num_tokens_from_string(
+        self, string: str, encoding_name="cl100k_base"
+    ) -> int:
+        """Returns the number of tokens in a text string."""
+        encoding = tiktoken.get_encoding(encoding_name)
+        num_tokens = len(encoding.encode(string))
+        return num_tokens
+
+    def transcript_splitter(self, chunk_size=3000, chunk_overlap=200):
+        markdown_splitter = MarkdownTextSplitter(
+            chunk_size=chunk_size, chunk_overlap=chunk_overlap
+        )
+        transcript_chunks = markdown_splitter.create_documents(
+            [self.raw_transcript]
+        )
+        return transcript_chunks
+
+    def transcript2insights(self, transcript):
+        system_template = "You are a helpful assistant that summarizes transcripts of podcasts or lectures."
+        system_prompt = SystemMessagePromptTemplate.from_template(
+            system_template
+        )
+        human_template = """From this chunk of a presentation transcript, extract a short list of key insights. \
+            Skip explaining what you're doing, labeling the insights and writing conclusion paragraphs. \
+            The insights have to be phrased as statements of facts with no references to the presentation or the transcript. \
+            Statements have to be full sentences and in terms of words and phrases as close as possible to those used in the transcript. \
+            Keep as much detail as possible. The transcript of the presentation is delimited in triple backticks.
+
+            Desired output format:
+            - [Key insight #1]
+            - [Key insight #2]
+            - [...]
+
+            Transcript:
+            ```{transcript}```"""
+        human_prompt = HumanMessagePromptTemplate.from_template(human_template)
+        chat_prompt = ChatPromptTemplate.from_messages(
+            [system_prompt, human_prompt]
+        )
+
+        result = self.chat(
+            chat_prompt.format_prompt(transcript=transcript).to_messages()
+        )
+
+        return result.content
+
+    def create_essay_insights(self, transcript_chunks, verbose=True):
+        response = ""
+        for i, text in enumerate(transcript_chunks):
+            insights = self.transcript2insights(text.page_content)
+            response = "\n".join([response, insights])
+            if verbose:
+                print(
+                    f"\nInsights extracted from chunk {i+1}/{len(transcript_chunks)}:\n{insights}"
+                )
+        return response
+
+    def create_blueprint(self, statements, verbose=True):
+        system_template = """You are a helpful AI blogger who writes essays on technical topics."""
+        system_prompt = SystemMessagePromptTemplate.from_template(
+            system_template
+        )
+
+        human_template = """Organize the following list of statements (delimited in triple backticks) to create the outline \
+            for a blog post in JSON format. The highest level is the most plausible statement as the overarching thesis \
+            statement of the post, the next layers are statements providing supporting arguments for the thesis statement. \
+            The last layer are pieces of evidence for each of the supporting arguments, directly quoted from the provided \
+            list of statements. Use as many of the provided statements as possible. Keep their wording as is without paraphrasing them. \
+            Retain as many technical details as possible. The thesis statement, supporting arguments, and evidences must be \
+            full sentences containing claims. Label each layer with the appropriate level title and create the desired JSON output format below. \
+            Only output the JSON and skip explaining what you're doing:
+
+            Desired output format:
+            {{
+            "Thesis Statement": "...",
+            "Supporting Arguments": [
+                {{
+                "Argument": "...",
+                "Evidence": [
+                    "...", "...", "...", ...
+                ]
+                }},
+                {{
+                "Argument": "...",
+                "Evidence": [
+                    "...", "...", "...", ...
+                ]
+                }},
+                ...
+            ]
+            }}
+
+            Statements:
+            ```{statements}```"""
+        human_prompt = HumanMessagePromptTemplate.from_template(human_template)
+        chat_prompt = ChatPromptTemplate.from_messages(
+            [system_prompt, human_prompt]
+        )
+
+        outline = self.chat(
+            chat_prompt.format_prompt(statements=statements).to_messages()
+        )
+
+        if verbose:
+            print(f"\nEssay outline: {outline.content}\n")
+        return outline.content
+
+    # @timer_decorator
+    def full_transcript2outline_json(self, verbose=True):
+        print("\nChunking transcript...")
+        transcript_docs = self.transcript_splitter()
+        t1 = time.time()
+        print("\nExtracting key insights...")
+        essay_insights = self.create_essay_insights(transcript_docs, verbose)
+        t2 = time.time() - t1
+        print("\nCreating essay outline...")
+        t1 = time.time()
+        blueprint = self.create_blueprint(essay_insights, verbose)
+        t3 = time.time() - t1
+        if verbose:
+            print()
+            print(f"Extracted essay insights in {t2:.2f} seconds.")
+            print(f"Created essay blueprint in {t3:.2f} seconds.")
+        return blueprint

demo/blog_writer/requirements.txt

@@ -0,0 +1,7 @@
+pdfminer.six
+sentence-transformers
+langchain==0.0.332
+python-dotenv>=1.0.0
+openai>=0.28.0
+tiktoken>=0.4.0
+sherpa-ai >= 0.2.1