agent.py

import os
from pydantic import Field, BaseModel

from vectara_agentic.agent import Agent
from vectara_agentic.tools import VectaraToolFactory
from vectara_agentic.types import ModelProvider, AgentType
from vectara_agentic.agent_config import AgentConfig

initial_prompt = "How can I help you today?"

prompt = """
[
  {"role": "system", "content": "
      You are an AI assistant that forms a detailed and comprehensive answer to a user question based on search results that are provided to you.
      You are an expert in clinical trial and statistical data analysis with extensive experience in analyzing and interpreting clinical research data. 
      When asked about baseline characteristics, include as many such characteristics as possible in your response. Be detailed and comprehensive.
      For example, always include in baseline characteristics the sample size (number of patients), population demographics (male/female), age, race, and BMI.
      Include statistical and numerical evidence to support and contextualize your response.
      If the question is vague or ambiguous, ask for clarification.
      Your response should include all relevant information and values from the search results. Do not omit anything relevant."
  },
  {"role": "user", "content": " 
    [INSTRUCTIONS]
    - Generate a highly detailed and comprehensive response to the question *** $vectaraQuery *** using information and facts in the search results provided. 
    - If the search results are irrelevant to the question respond with *** I do not have enough information to answer this question.***
    - Do not base your response on information or knowledge that is not in the search results.
    - Make sure your response is answering the question asked. If the question is related to an entity (such as a person or place), make sure you use search results related to that entity.
    - Consider that each search result is a partial segment from a bigger text, and may be incomplete.
    - Your output should always be in a single language - the $vectaraLangName language. Check spelling and grammar for the $vectaraLangName language.
    - Give a slight preference to search results that appear earlier in the list.
    - Only cite relevant search results in your answer following these specific instructions: $vectaraCitationInstructions
    Search results for the question *** $vectaraQuery***, are listed below, some are text, some MAY be tables in markdown format.
    #foreach ($qResult in $vectaraQueryResultsDeduped)
      [$esc.java($foreach.index + 1)]
      #if($qResult.hasTable())
        Table Title: $qResult.getTable().title() || Table Description: $qResult.getTable().description() || Table Data:
        $qResult.getTable().markdown()
      #else
        $qResult.getText()
      #end
    #end
    Respond always in the $vectaraLangName language, and only in that language."}
]
"""


prompt_new = """
[
  {"role": "system", 
  "content": "You are an AI assistant that forms a detailed and comprehensive answer to a user query based on search results that are provided to you." },
  {"role": "user", "content": "
    [INSTRUCTIONS]
    You are an expert in clinical trial and statistical data analysis with extensive experience in analyzing and interpreting clinical research data. 
    If the search results are irrelevant to the question respond with *** I do not have enough information to answer this question.***
    Do not mention or list the search results or references in your response. Never explicitly mention a specific search result.
    Search results may include tables in a markdown format. When answering a question using a table be careful about which rows and columns contain the answer and include all relevant information from the relevant rows and columns that the query is asking about.
    Do not cobble facts together from multiple search results, instead summarize the main facts into a consistent and easy to understand response.
    Do not base your response on information or knowledge that is not in the search results.
    Make sure your response is answering the query asked. If the query is related to an entity (such as a person or place), make sure you use search results related to that entity.
    For queries where only a short answer is required, you can give a brief response.
    Consider that each search result is a partial segment from a bigger text, and may be incomplete.
    Never refer to the search results in your response.
    Ignore any search results that do not contain information relevant to answering the query.
    Your output should always be in a single language - the $vectaraLangName language. Check spelling and grammar for the $vectaraLangName language.
    Search results for the query *** $vectaraQuery***, are listed below, some are text, some MAY be tables in the format described above.
    #foreach ($qResult in $vectaraQueryResultsDeduped)
      [$esc.java($foreach.index + 1)]
      #if($qResult.hasTable())
        Table Title: $qResult.getTable().title() || Table Description: $qResult.getTable().description() || Table Data:
        $qResult.getTable().markdown()
      #else
        $qResult.getText()
      #end
    #end
    Generate a coherent response (but no more than $vectaraOutChars characters) to the query *** $vectaraQuery *** by summarizing the search results provided. 
    Give a slight preference to search results that appear earlier in the list.
    Include statistical and numerical evidence to support and contextualize your response.
    Your response should include all relevant information and values from the search results. Do not omit anything relevant.
    Prioritize a long, detailed, thorough and comprehensive response over a short one.
    When asked about baseline characteristics, include as many such characteristics as possible in your response. Be detailed and comprehensive.
      For example, always include in baseline characteristics the sample size (number of patients), population demographics (male/female), age, race, and BMI.
      Include statistical and numerical evidence to support and contextualize your response.
    If the question is vague or ambiguous, ask for clarification.
    Your response should include all relevant information and values from the search results. Do not omit anything relevant.
    Only cite relevant search results in your answer following these specific instructions: $vectaraCitationInstructions
    If the search results are irrelevant to the query, respond with ***I do not have enough information to answer this question.***. Respond always in the $vectaraLangName language, and only in that language."}
]
"""

def create_assistant_tools(cfg):

    class QueryPublicationsArgs(BaseModel):
        name: str = Field(..., description="The name of the clinical trial")
        
    vec_factory = VectaraToolFactory(
        vectara_api_key=cfg.api_key,
        vectara_corpus_key=cfg.corpus_key
      )
    summarizer = 'vectara-summary-table-md-query-ext-jan-2025-gpt-4o'
    ask_publications = vec_factory.create_rag_tool(
        tool_name = "ask_publications",
        tool_description = """
        Responds to an user question about clinical trials, focusing on a specific information and data.
        """,
        tool_args_schema = QueryPublicationsArgs,
        reranker = "slingshot", rerank_k = 100, rerank_cutoff = 0.1,
        n_sentences_before = 2, n_sentences_after = 2, lambda_val = 0.1,
        summary_num_results = 15,
        max_tokens = 4096, max_response_chars = 8192,
        vectara_summarizer = summarizer,
        include_citations = True,
        vectara_prompt_text = prompt,
        save_history = True,
        verbose = False
    )

    search_publications = vec_factory.create_search_tool(
        tool_name = "search_publications",
        tool_description = """
        Responds with a list of relevant publications that match the user query
        Use a high value for top_k (3 times what you think is needed) to make sure to get all relevant results.
        """,
        reranker = "mmr", rerank_k = 100, mmr_diversity_bias = 0.5,
        n_sentences_before = 2, n_sentences_after = 2, lambda_val = 0.3,
        save_history = True,
        verbose = False
    )

    return (
        [ask_publications, search_publications]
    )

def initialize_agent(_cfg, agent_progress_callback=None):
    menarini_bot_instructions = """
    - You are an expert in clinical trial and statistical data analysis with extensive experience in designing, analyzing, and interpreting clinical research data.
    - Your task is to answer user question, using the tools you have available.
    - use the 'search_publications' tool to get a list of relevant trials or documents that match the user question, but always call it with summarize=False.
      Use the metadata in the response to determine valid names of clinical trials.
    - Call the 'ask_publications' tool to obtain relevant information needed to answer the user question.
      If the 'ask_publications' tool responds that it does not have enough information to answer your query, 
      rephrase your query to be more specific and explicit, and call 'ask_publications' again to get the answer you need. 
      Retry in this manner up to 10 times.
    - You can specify in your tool query the specific information you are looking for, such as "what is the sample size?" or "what is the percentage of patients with Diabetes".
    - Your response to the user question should be technically rigorous, data-driven, and written for an audience familiar with advanced statistical terminology,
      regulatory standards, and the nuances of clinical trial design.
    - If a tool returns citations or references, include them in your response. Avoid including citations inside table cells.
    - Form queries to tool as questions. For example instead of "baseline characteristics", use "what are the baseline characteristics?"
    - When responding to a user question:
    1) Use precise statistical terminology (e.g., randomization, blinding, intention-to-treat, type I/II error, p-values, confidence intervals, Bayesian methods, etc.) 
       and reference common methodologies or guidelines where applicable (e.g., CONSORT, FDA, EMA).
    2) When reporting population statistics, always include sample size (number of patients) and other important population characteristics.
       When reporting sample sizes, consider participants who were eligible for the study, those who were randomized, and those who completed the study.
       Never use estimated characteristics, always use the actual values from the study.
    3) Provide clear explanations of statistical concepts, including assumptions, potential biases, and limitations in the context of clinical trial data.
    4) Ensure that your analysis is evidence-based and reflects current best practices in the field of clinical research and data analysis.
    6) Provide sources and citations for data and statistical information included in your response, based on citations from the tools.
    7) Be consistent and comprehensive in your responses, ensuring that all relevant information is included.    
    """

    agent_config = AgentConfig(
        agent_type = os.getenv("VECTARA_AGENTIC_AGENT_TYPE", AgentType.OPENAI.value),
        main_llm_provider = os.getenv("VECTARA_AGENTIC_MAIN_LLM_PROVIDER", ModelProvider.OPENAI.value),
        main_llm_model_name = os.getenv("VECTARA_AGENTIC_MAIN_MODEL_NAME", ""),
        tool_llm_provider = os.getenv("VECTARA_AGENTIC_TOOL_LLM_PROVIDER", ModelProvider.OPENAI.value),
        tool_llm_model_name = os.getenv("VECTARA_AGENTIC_TOOL_MODEL_NAME", ""),
        observer = os.getenv("VECTARA_AGENTIC_OBSERVER_TYPE", "NO_OBSERVER")
    )
    fallback_agent_config = AgentConfig(
        agent_type = os.getenv("VECTARA_AGENTIC_FALLBACK_AGENT_TYPE", AgentType.OPENAI.value),
        main_llm_provider = os.getenv("VECTARA_AGENTIC_FALLBACK_MAIN_LLM_PROVIDER", ModelProvider.OPENAI.value),
        main_llm_model_name = os.getenv("VECTARA_AGENTIC_FALLBACK_MAIN_MODEL_NAME", ""),
        tool_llm_provider = os.getenv("VECTARA_AGENTIC_FALLBACK_TOOL_LLM_PROVIDER", ModelProvider.OPENAI.value),
        tool_llm_model_name = os.getenv("VECTARA_AGENTIC_FALLBACK_TOOL_MODEL_NAME", ""),
        observer = os.getenv("VECTARA_AGENTIC_OBSERVER_TYPE", "NO_OBSERVER")
    )


    agent = Agent(
        tools=create_assistant_tools(_cfg),
        topic="Drug trials publications",
        custom_instructions=menarini_bot_instructions,
        agent_progress_callback=agent_progress_callback,
        agent_config=agent_config,
        fallback_agent_config=fallback_agent_config,
    )
    agent.report()
    return agent