app.py

import gradio as gr
import os
import time
import json
import re
from uuid import uuid4
from datetime import datetime
from duckduckgo_search import DDGS
from sentence_transformers import SentenceTransformer, util
from typing import List, Dict, Any, Optional, Union, Tuple
import logging
import numpy as np
from collections import deque
from huggingface_hub import InferenceClient
import requests
import arxiv
import scholarly
import pymed
import wikipedia
import trafilatura
from trafilatura import extract, fetch_url
import pickle
import faiss
import threading
from concurrent.futures import ThreadPoolExecutor, as_completed
import tiktoken

logging.basicConfig(level=logging.INFO, format='%(asctime)s - %(name)s - %(levelname)s - %(message)s')
logger = logging.getLogger(__name__)

HF_API_KEY = os.environ.get("HF_API_KEY")
if not HF_API_KEY:
    raise ValueError("Please set the HF_API_KEY environment variable.")

client = InferenceClient(provider="hf-inference", api_key=HF_API_KEY)

MAIN_LLM_MODEL = "mistralai/Mistral-Nemo-Instruct-2407"
REASONING_LLM_MODEL = "deepseek-ai/DeepSeek-R1-Distill-Qwen-32B"
CRITIC_LLM_MODEL = "Qwen/QwQ-32B-Preview"
ENSEMBLE_MODELS = [MAIN_LLM_MODEL, REASONING_LLM_MODEL, CRITIC_LLM_MODEL]  # Keep, but expand upon.

MAX_ITERATIONS = 40  # Increased for deeper research.
TIMEOUT = 180 # Longer timeout for larger models / complex tasks.
RETRY_DELAY = 10 # longer delay
NUM_RESULTS = 20
SIMILARITY_THRESHOLD = 0.15
MAX_CONTEXT_ITEMS = 50  # Increased context window.
MAX_HISTORY_ITEMS = 12
MAX_FULL_TEXT_LENGTH = 20000 # larger document size
FAISS_INDEX_PATH = "research_index.faiss"
RESEARCH_DATA_PATH = "research_data.pkl"
PAPER_SUMMARIES_PATH = "paper_summaries.pkl" #New path for storing paper summary


try:
    main_similarity_model = SentenceTransformer('sentence-transformers/all-mpnet-base-v2')
    concept_similarity_model = SentenceTransformer('sentence-transformers/all-MiniLM-L6-v2')
    document_similarity_model = SentenceTransformer('sentence-transformers/multi-qa-mpnet-base-dot-v1')

    embedding_dim = document_similarity_model.get_sentence_embedding_dimension()
    if os.path.exists(FAISS_INDEX_PATH):
        index = faiss.read_index(FAISS_INDEX_PATH)
        logger.info(f"Loaded FAISS index from {FAISS_INDEX_PATH}")
    else:
        index = faiss.IndexFlatIP(embedding_dim)  # Use IndexFlatIP for inner product (cosine similarity).
        logger.info("Created a new FAISS index.")
except Exception as e:
    logger.error(f"Failed to load models or initialize FAISS: {e}")
    raise

def get_token_count(text):
    try:
        encoding = tiktoken.get_encoding("cl100k_base")
        return len(encoding.encode(text))
    except:
        return len(text.split()) * 1.3

def save_research_data(data, index):
    try:
        with open(RESEARCH_DATA_PATH, "wb") as f:
            pickle.dump(data, f)
        faiss.write_index(index, FAISS_INDEX_PATH)
        logger.info(f"Research data and index saved to {RESEARCH_DATA_PATH} and {FAISS_INDEX_PATH}")
    except Exception as e:
        logger.error(f"Error saving research data: {e}")

def load_research_data():
    if os.path.exists(RESEARCH_DATA_PATH):
        try:
            with open(RESEARCH_DATA_PATH, "rb") as f:
                data = pickle.load(f)
                logger.info(f"Loaded research data from {RESEARCH_DATA_PATH}")
                return data
        except Exception as e:
            logger.error(f"Error loading research data: {e}")
            return {}
    else:
        logger.info("No existing research data found.")
        return {}

def save_paper_summaries(summaries: Dict[str, str]):
    try:
        with open(PAPER_SUMMARIES_PATH, "wb") as f:
            pickle.dump(summaries, f)
        logger.info(f"Paper summaries saved to {PAPER_SUMMARIES_PATH}")
    except Exception as e:
        logger.error(f"Error saving paper summaries: {e}")

def load_paper_summaries() -> Dict[str, str]:
    if os.path.exists(PAPER_SUMMARIES_PATH):
        try:
            with open(PAPER_SUMMARIES_PATH, "rb") as f:
                data = pickle.load(f)
                logger.info(f"Loaded paper summaries from {PAPER_SUMMARIES_PATH}")
                return data
        except Exception as e:
            logger.error(f"Error loading paper summaries: {e}")
            return {}
    else:
        logger.info("No existing paper summaries found.")
        return {}


def hf_inference(model_name, prompt, max_tokens=2000, retries=5):
    for attempt in range(retries):
        try:
            messages = [{"role": "user", "content": prompt}]
            response = client.chat.completions.create(
                model=model_name,
                messages=messages,
                max_tokens=max_tokens
            )
            return {"generated_text": response.choices[0].message.content}
        except Exception as e:
            if attempt == retries - 1:
                logger.error(f"Request failed after {retries} retries: {e}")
                return {"error": f"Request failed after {retries} retries: {e}"}
            time.sleep(RETRY_DELAY * (1 + attempt))
    return {"error": "Request failed after multiple retries."}

def ensemble_inference(prompt, models=ENSEMBLE_MODELS, max_tokens=1500):
    results = []
    with ThreadPoolExecutor(max_workers=len(models)) as executor:
        future_to_model = {executor.submit(hf_inference, model, prompt, max_tokens): model for model in models}
        for future in as_completed(future_to_model):
            model = future_to_model[future]
            try:
                result = future.result()
                if "generated_text" in result:
                    results.append({"model": model, "text": result["generated_text"]})
            except Exception as e:
                logger.error(f"Error with model {model}: {e}")

    if not results:
        return {"error": "All models failed to generate responses"}

    if len(results) == 1:
        return {"generated_text": results[0]["text"]}

    synthesis_prompt = "Synthesize these expert responses into a single coherent answer:\n\n"
    for result in results:
        synthesis_prompt += f"Expert {results.index(result) + 1} ({result['model'].split('/')[-1]}):\n{result['text']}\n\n"

    synthesis = hf_inference(MAIN_LLM_MODEL, synthesis_prompt) # Use a consistent model for final synthesis
    if "generated_text" in synthesis:
        return synthesis
    else:
        return {"generated_text": max(results, key=lambda x: len(x["text"]))["text"]} # Fallback

def tool_search_web(query: str, num_results: int = NUM_RESULTS, safesearch: str = "moderate",
                   time_filter: Optional[str] = None, region: str = "wt-wt", language: str = "en-us") -> list:
    try:
        with DDGS() as ddgs:
            kwargs = {
                "keywords": query,
                "max_results": num_results,
                "safesearch": safesearch,
                "region": region,
                "hreflang": language,
            }
            if time_filter:
                if time_filter in ['d', 'w', 'm', 'y']:
                    kwargs["time"] = time_filter

            results = [r for r in ddgs.text(**kwargs)]
            if results:
                return [{"title": r["title"], "snippet": r["body"], "url": r["href"]} for r in results]
            else:
                if time_filter and "time" in kwargs:
                    del kwargs["time"]
                    results = [r for r in ddgs.text(**kwargs)]
                    if results:
                        return [{"title": r["title"], "snippet": r["body"], "url": r["href"]} for r in results]
                return []
    except Exception as e:
        logger.error(f"DuckDuckGo search error: {e}")
        return []

def tool_search_arxiv(query: str, max_results: int = 5) -> list:
    try:
        client = arxiv.Client()
        search = arxiv.Search(
            query=query,
            max_results=max_results,
            sort_by=arxiv.SortCriterion.Relevance
        )
        results = []
        for paper in client.results(search):
            results.append({
                "title": paper.title,
                "snippet": paper.summary[:500] + "..." if len(paper.summary) > 500 else paper.summary,
                "url": paper.pdf_url,
                "authors": ", ".join(author.name for author in paper.authors),
                "published": paper.published.strftime("%Y-%m-%d") if paper.published else "Unknown",
                "source": "arXiv"
            })
        return results
    except Exception as e:
        logger.error(f"arXiv search error: {e}")
        return []

def tool_search_pubmed(query: str, max_results: int = 5) -> list:
    try:
        pubmed = pymed.PubMed(tool="ResearchAssistant", email="[email protected]")
        results = list(pubmed.query(query, max_results=max_results))

        output = []
        for article in results:
            try:
                data = article.toDict()
                output.append({
                    "title": data.get("title", "No title"),
                    "snippet": data.get("abstract", "No abstract")[:500] + "..." if data.get("abstract", "") and len(data.get("abstract", "")) > 500 else data.get("abstract", "No abstract"),
                    "url": f"https://pubmed.ncbi.nlm.nih.gov/{data.get('pubmed_id')}/",
                    "authors": ", ".join(author.get("name", "") for author in data.get("authors", [])),
                    "published": data.get("publication_date", "Unknown"),
                    "source": "PubMed"
                })
            except:
                continue
        return output
    except Exception as e:
        logger.error(f"PubMed search error: {e}")
        return []

def tool_search_wikipedia(query: str, max_results: int = 3) -> list:
    try:
        search_results = wikipedia.search(query, results=max_results)
        results = []

        for title in search_results:
            try:
                page = wikipedia.page(title)
                summary = page.summary
                snippet = summary[:500] + "..." if len(summary) > 500 else summary
                results.append({
                    "title": page.title,
                    "snippet": snippet,
                    "url": page.url,
                    "source": "Wikipedia"
                })
            except (wikipedia.exceptions.DisambiguationError, wikipedia.exceptions.PageError):
                continue

        return results
    except Exception as e:
        logger.error(f"Wikipedia search error: {e}")
        return []
    
def tool_search_scholar(query: str, max_results: int = 5) -> list:
    try:
        search_query = scholarly.search_pubs(query)
        results = []
        for _ in range(max_results):
            try:
                result = next(search_query)
                results.append({
                    "title": result.get("bib", {}).get("title", "No title"),
                    "snippet": result.get("bib", {}).get("abstract", "No abstract")[:500] + "..." if result.get("bib", {}).get("abstract") else result.get("bib", {}).get("abstract", "No abstract"),
                    "url": result.get("pub_url", "#"),
                    "authors": ", ".join(result.get("bib", {}).get("author", [])),
                    "published": result.get("bib", {}).get("pub_year", "Unknown"),
                    "source": "Google Scholar"
                })
            except StopIteration:
                break
            except Exception as e:
                logger.warning(f"Error processing Scholar result: {e}")
                continue
        return results
    except Exception as e:
        logger.error(f"Google Scholar search error: {e}")
        return []

def extract_article_content(url: str) -> str:
    try:
        downloaded = fetch_url(url)
        if downloaded is None:
            return ""
        return extract(downloaded, favor_precision=True)
    except Exception as e:
        logger.error(f"Failed to extract article content from {url}: {e}")
        return ""

def tool_reason(prompt: str, search_results: list, reasoning_context: list = [],
               critique: str = "", focus_areas: list = []) -> str:
    if not search_results:
        return "No search results to reason about."

    reasoning_input = "Reason about the following search results in relation to the prompt:\n\n"
    reasoning_input += f"Prompt: {prompt}\n\n"

    if focus_areas:
        reasoning_input += f"Focus particularly on these aspects: {', '.join(focus_areas)}\n\n"

    results_by_source = {}
    for i, result in enumerate(search_results):
        source = result.get('source', 'Web Search') # Default to 'Web Search'
        if source not in results_by_source:
            results_by_source[source] = []
        results_by_source[source].append((i, result))

    for source, results in results_by_source.items():
        reasoning_input += f"\n--- {source} Results ---\n"
        for i, result in results:
            reasoning_input += f"- Result {i + 1}: Title: {result['title']}\n  Snippet: {result['snippet']}\n"
            if 'authors' in result:
                reasoning_input += f"  Authors: {result['authors']}\n"
            if 'published' in result:
                reasoning_input += f"  Published: {result['published']}\n"
            reasoning_input += "\n"

    if reasoning_context:
        recent_context = reasoning_context[-MAX_HISTORY_ITEMS:]  # Limit history
        reasoning_input += "\nPrevious Reasoning Context:\n" + "\n".join(recent_context)

    if critique:
        reasoning_input += f"\n\nRecent critique to address: {critique}\n"

    reasoning_input += "\nProvide a thorough, nuanced analysis that builds upon previous reasoning if applicable. Consider multiple perspectives, potential contradictions in the search results, and the reliability of different sources.  Address any specific critiques."

    reasoning_output = ensemble_inference(reasoning_input) # Use ensemble for high-quality reasoning.

    if isinstance(reasoning_output, dict) and "generated_text" in reasoning_output:
        return reasoning_output["generated_text"].strip()
    else:
        logger.error(f"Failed to generate reasoning: {reasoning_output}")
        return "Could not generate reasoning due to an error."

def tool_summarize(insights: list, prompt: str, contradictions: list = []) -> str:
    if not insights:
        return "No insights to summarize."

    summarization_input = f"Synthesize the following insights into a cohesive and comprehensive summary regarding: '{prompt}'\n\n"

    max_tokens = 12000  # Increased token limit
    selected_insights = []
    token_count = get_token_count(summarization_input) + get_token_count("\n\n".join(contradictions))

    for insight in reversed(insights):
        insight_tokens = get_token_count(insight)
        if token_count + insight_tokens < max_tokens:
            selected_insights.insert(0, insight)
            token_count += insight_tokens
        else:
            break

    summarization_input += "\n\n".join(selected_insights)

    if contradictions:
        summarization_input += "\n\nAddress these specific contradictions:\n" + "\n".join(contradictions)

    summarization_input += "\n\nProvide a well-structured summary that:\n1. Presents the main findings\n2. Acknowledges limitations and uncertainties\n3. Highlights areas of consensus and disagreement\n4. Suggests potential directions for further inquiry\n5. Evaluates the strength of evidence for key claims"

    summarization_output = ensemble_inference(summarization_input)

    if isinstance(summarization_output, dict) and "generated_text" in summarization_output:
        return summarization_output["generated_text"].strip()
    else:
        logger.error(f"Failed to generate summary: {summarization_output}")
        return "Could not generate a summary due to an error."

def tool_generate_search_query(prompt: str, previous_queries: list = [],
                              failed_queries: list = [], focus_areas: list = []) -> str:
    query_gen_input = f"Generate an effective search query for the following prompt: {prompt}\n"

    if previous_queries:
        recent_queries = previous_queries[-MAX_HISTORY_ITEMS:]
        query_gen_input += "Previous search queries:\n" + "\n".join(recent_queries) + "\n"

    if failed_queries:
        query_gen_input += "These queries didn't yield useful results:\n" + "\n".join(failed_queries) + "\n"

    if focus_areas:
        query_gen_input += f"Focus particularly on these aspects: {', '.join(focus_areas)}\n"

    query_gen_input += "Refine the search query based on previous queries, aiming for more precise results. Consider using advanced search operators like site:, filetype:, intitle:, etc. when appropriate. Make sure the query is well-formed for academic and scientific search engines.\n"
    query_gen_input += "Search Query:"

    query_gen_output = hf_inference(MAIN_LLM_MODEL, query_gen_input)

    if isinstance(query_gen_output, dict) and 'generated_text' in query_gen_output:
        return query_gen_output['generated_text'].strip()

    logger.error(f"Failed to generate search query: {query_gen_output}")
    return ""

def tool_critique_reasoning(reasoning_output: str, prompt: str,
                           previous_critiques: list = []) -> str:
    critique_input = f"Critically evaluate the following reasoning output in relation to the prompt:\n\nPrompt: {prompt}\n\nReasoning: {reasoning_output}\n\n"

    if previous_critiques:
        critique_input += "Previous critiques that should be addressed:\n" + "\n".join(previous_critiques[-MAX_HISTORY_ITEMS:]) + "\n\n"

    critique_input += "Identify any flaws, biases, logical fallacies, unsupported claims, or areas for improvement. Be specific and constructive. Suggest concrete ways to enhance the reasoning. Also evaluate the strength of evidence and whether conclusions are proportionate to the available information."

    critique_output = hf_inference(CRITIC_LLM_MODEL, critique_input) # Use specialized critique model.

    if isinstance(critique_output, dict) and "generated_text" in critique_output:
        return critique_output["generated_text"].strip()

    logger.error(f"Failed to generate critique: {critique_output}")
    return "Could not generate a critique due to an error."

def tool_identify_contradictions(insights: list) -> list:
    if len(insights) < 2:
        return []

    max_tokens = 12000  # Increased token limit for potentially more contradictions
    selected_insights = []
    token_count = 0

    for insight in reversed(insights):
        insight_tokens = get_token_count(insight)
        if token_count + insight_tokens < max_tokens:
            selected_insights.insert(0, insight)
            token_count += insight_tokens
        else:
            break

    contradiction_input = "Identify specific contradictions in these insights:\n\n" + "\n\n".join(selected_insights)
    contradiction_input += "\n\nList each contradiction as a separate numbered point. For each contradiction, cite the specific claims that are in tension and evaluate which claim is better supported. If no contradictions exist, respond with 'No contradictions found.'"

    contradiction_output = hf_inference(CRITIC_LLM_MODEL, contradiction_input)  # Use critique model

    if isinstance(contradiction_output, dict) and "generated_text" in contradiction_output:
        result = contradiction_output["generated_text"].strip()
        if result == "No contradictions found.":
            return []
        # More robust contradiction extraction, handles multi-sentence contradictions
        contradictions = re.findall(r'\d+\.\s+(.*?)(?=\d+\.|$)', result, re.DOTALL)
        return [c.strip() for c in contradictions if c.strip()]

    logger.error(f"Failed to identify contradictions: {contradiction_output}")
    return []

def tool_identify_focus_areas(prompt: str, insights: list = [],
                             failed_areas: list = []) -> list:
    focus_input = f"Based on this research prompt: '{prompt}'\n\n"

    if insights:
        recent_insights = insights[-5:] if len(insights) > 5 else insights
        focus_input += "And these existing insights:\n" + "\n".join(recent_insights) + "\n\n"

    if failed_areas:
        focus_input += f"These focus areas didn't yield useful results: {', '.join(failed_areas)}\n\n"

    focus_input += "Identify 3-5 specific aspects that should be investigated further to get a complete understanding. Be precise and prioritize underexplored areas. For each suggested area, briefly explain why it's important to investigate."

    focus_output = hf_inference(MAIN_LLM_MODEL, focus_input)  # Consistent model

    if isinstance(focus_output, dict) and "generated_text" in focus_output:
        result = focus_output["generated_text"].strip()
        # More robust extraction, handles different list formats
        areas = re.findall(r'(?:^|\n)(?:\d+\.|\*|\-)\s*(.*?)(?=(?:\n(?:\d+\.|\*|\-|$))|$)', result)
        return [area.strip() for area in areas if area.strip()][:5]

    logger.error(f"Failed to identify focus areas: {focus_output}")
    return []

def add_to_faiss_index(text: str):
    embedding = document_similarity_model.encode(text, convert_to_tensor=True)
    embedding_np = embedding.cpu().numpy().reshape(1, -1)
    if embedding_np.shape[1] != embedding_dim:
        logger.error(f"Embedding dimension mismatch: expected {embedding_dim}, got {embedding_np.shape[1]}")
        return
    faiss.normalize_L2(embedding_np)  # Normalize for cosine similarity.
    index.add(embedding_np)

def search_faiss_index(query: str, top_k: int = 5) -> List[str]:
    query_embedding = document_similarity_model.encode(query, convert_to_tensor=True)
    query_embedding_np = query_embedding.cpu().numpy().reshape(1, -1)
    faiss.normalize_L2(query_embedding_np)
    distances, indices = index.search(query_embedding_np, top_k)
    return indices[0].tolist()

def filter_results(search_results, prompt, previous_snippets=None):
    if not main_similarity_model or not search_results:
        return search_results

    try:
        prompt_embedding = main_similarity_model.encode(prompt, convert_to_tensor=True)
        filtered_results = []

        seen_snippets = set()
        if previous_snippets:
            seen_snippets.update(previous_snippets)

        for result in search_results:
            combined_text = result['title'] + " " + result['snippet']

            if result['snippet'] in seen_snippets:  # Prevent exact duplicates
                continue

            result_embedding = main_similarity_model.encode(combined_text, convert_to_tensor=True)
            cosine_score = util.pytorch_cos_sim(prompt_embedding, result_embedding)[0][0].item()
            
            if cosine_score >= SIMILARITY_THRESHOLD:
                result['relevance_score'] = cosine_score
                filtered_results.append(result)
                seen_snippets.add(result['snippet']) # Add snippets after filtering
                add_to_faiss_index(result['snippet'])


        filtered_results.sort(key=lambda x: x.get('relevance_score', 0), reverse=True)  # Sort by relevance.
        return filtered_results

    except Exception as e:
        logger.error(f"Error during filtering: {e}")
        return search_results  # Return original results on error.

def tool_extract_key_entities(prompt: str) -> list:
    entity_input = f"Extract the key entities (people, organizations, concepts, technologies, events, time periods, locations, etc.) from this research prompt that should be investigated individually:\n\n{prompt}\n\nList the 5-7 most important entities, one per line, with a brief explanation (2-3 sentences) of why each is central to the research question."

    entity_output = hf_inference(MAIN_LLM_MODEL, entity_input)

    if isinstance(entity_output, dict) and "generated_text" in entity_output:
        result = entity_output["generated_text"].strip()
        entities = [e.strip() for e in result.split('\n') if e.strip()]
        return entities[:7] # Limit to top 7 entities

    logger.error(f"Failed to extract key entities: {entity_output}")
    return []

def tool_meta_analyze(entity_insights: Dict[str, list], prompt: str) -> str:
    if not entity_insights:
        return "No entity insights to analyze."

    meta_input = f"Perform a meta-analysis across these different entities related to the prompt: '{prompt}'\n\n"

    for entity, insights in entity_insights.items():
        if insights:
            meta_input += f"\n--- {entity} ---\n" + insights[-1] + "\n"  # Most recent insight for each entity

    meta_input += "\nProvide a high-level synthesis that identifies:\n1. Common themes across entities\n2. Important differences and contradictions\n3. How these entities interact or influence each other\n4. The broader implications for the original research question\n5. A systems-level understanding of how these elements fit together"

    meta_output = ensemble_inference(meta_input)  # Ensemble for meta-analysis

    if isinstance(meta_output, dict) and "generated_text" in meta_output:
        return meta_output["generated_text"].strip()

    logger.error(f"Failed to perform meta-analysis: {meta_output}")
    return "Could not generate a meta-analysis due to an error."

def tool_draft_research_plan(prompt: str, entities: list, focus_areas: list = []) -> str:
    plan_input = f"Create a detailed research plan for investigating this question: '{prompt}'\n\n"

    if entities:
        plan_input += "Key entities to investigate:\n" + "\n".join(entities) + "\n\n"

    if focus_areas:
        plan_input += "Additional focus areas:\n" + "\n".join(focus_areas) + "\n\n"

    plan_input += "The research plan should include:\n"
    plan_input += "1. Main research questions and sub-questions\n"
    plan_input += "2. Methodology for investigating each aspect\n"
    plan_input += "3. Potential sources and databases to consult\n"
    plan_input += "4. Suggested sequence of investigation\n"
    plan_input += "5. Potential challenges and how to address them\n"
    plan_input += "6. Criteria for evaluating the quality of findings"

    plan_output = hf_inference(REASONING_LLM_MODEL, plan_input) # Use reasoning model

    if isinstance(plan_output, dict) and "generated_text" in plan_output:
        return plan_output["generated_text"].strip()

    logger.error(f"Failed to generate research plan: {plan_output}")
    return "Could not generate a research plan due to an error."

def tool_extract_article(url: str) -> str:
    extracted_text = extract_article_content(url)
    return extracted_text if extracted_text else f"Could not extract content from {url}"

# New tool for summarizing a single paper
def tool_summarize_paper(paper_text: str) -> str:
  summarization_prompt = f"""Summarize this academic paper, focusing on the following:

1.  **Main Research Question(s):** What questions does the paper address?
2.  **Methodology:** Briefly describe the methods used (e.g., experiments, surveys, simulations, theoretical analysis).
3.  **Key Findings:** What are the most important results or conclusions?
4.  **Limitations:**  What are the acknowledged limitations of the study?
5.  **Implications:** What are the broader implications of the findings, according to the authors?

Paper Text:
{paper_text[:MAX_FULL_TEXT_LENGTH]}
"""  # Truncate if necessary
  summary = hf_inference(REASONING_LLM_MODEL, summarization_prompt, max_tokens=500)

  if isinstance(summary, dict) and "generated_text" in summary:
      return summary["generated_text"].strip()
  else:
      logger.error(f"Failed to generate summary: {summary}")
      return "Could not generate a summary due to an error."


tools = {
    "search_web": {
        "function": tool_search_web,
        "description": "Searches the web for information.",
        "parameters": {
            "query": {"type": "string", "description": "The search query."},
            "num_results": {"type": "integer", "description": "Number of results to return."},
            "time_filter": {"type": "string", "description": "Optional time filter (d, w, m, y)."},
            "region": {"type": "string", "description": "Optional region code."},
            "language": {"type": "string", "description": "Optional language code."}
        },
    },
    "search_arxiv": {
        "function": tool_search_arxiv,
        "description": "Searches arXiv for scientific papers.",
        "parameters": {
            "query": {"type": "string", "description": "The search query for scientific papers."},
            "max_results": {"type": "integer", "description": "Maximum number of papers to return."}
        },
    },
    "search_pubmed": {
        "function": tool_search_pubmed,
        "description": "Searches PubMed for medical and scientific literature.",
        "parameters": {
            "query": {"type": "string", "description": "The search query for medical literature."},
            "max_results": {"type": "integer", "description": "Maximum number of articles to return."}
        },
    },
    "search_wikipedia": {
        "function": tool_search_wikipedia,
        "description": "Searches Wikipedia for information.",
        "parameters": {
            "query": {"type": "string", "description": "The search query for Wikipedia."},
            "max_results": {"type": "integer", "description": "Maximum number of articles to return."}
        },
    },
    "search_scholar": {
        "function": tool_search_scholar,
        "description": "Searches Google Scholar for academic publications.",
        "parameters": {
            "query": {"type": "string", "description": "The search query for Google Scholar."},
            "max_results": {"type": "integer", "description": "Maximum number of articles to return."}
        }
    },
    "extract_article": {
        "function": tool_extract_article,
        "description": "Extracts the main content from a web article URL",
        "parameters": {
            "url": {"type": "string", "description": "The URL of the article to extract"}
        },
    },
        "summarize_paper": {
        "function": tool_summarize_paper,
        "description": "Summarizes the content of an academic paper.",
        "parameters": {
            "paper_text": {"type": "string", "description": "The full text of the paper to be summarized."},
        },
    },
    "reason": {
        "function": tool_reason,
        "description": "Analyzes and reasons about information.",
        "parameters": {
            "prompt": {"type": "string", "description": "The original prompt."},
            "search_results": {"type": "array", "description": "Search results to analyze."},
            "reasoning_context": {"type": "array", "description": "Previous reasoning outputs."},
            "critique": {"type": "string", "description": "Recent critique to address."},
            "focus_areas": {"type": "array", "description": "Specific aspects to focus on."}
        },
    },
    "summarize": {
        "function": tool_summarize,
        "description": "Synthesizes insights into a cohesive summary.",
        "parameters": {
            "insights": {"type": "array", "description": "Insights to summarize."},
            "prompt": {"type": "string", "description": "The original research prompt."},
            "contradictions": {"type": "array", "description": "Specific contradictions to address."}
        },
    },
    "generate_search_query": {
        "function": tool_generate_search_query,
        "description": "Generates an optimized search query",
        "parameters":{
            "prompt": {"type": "string", "description": "The original user prompt."},
            "previous_queries": {"type": "array", "description": "Previously used search queries."},
            "failed_queries": {"type": "array", "description": "Queries that didn't yield good results."},
            "focus_areas": {"type": "array", "description": "Specific aspects to focus on."}
        }
    },
    "critique_reasoning": {
        "function": tool_critique_reasoning,
        "description": "Critically evaluates reasoning output.",
        "parameters": {
            "reasoning_output": {"type": "string", "description": "The reasoning output to critique."},
            "prompt": {"type": "string", "description": "The original prompt."},
            "previous_critiques": {"type": "array", "description": "Previous critique outputs."}
        },
    },
    "identify_contradictions": {
        "function": tool_identify_contradictions,
        "description": "Identifies contradictions across multiple insights.",
        "parameters": {
            "insights": {"type": "array", "description": "Collection of insights to analyze for contradictions."},
                },
    },
    "identify_focus_areas": {
        "function": tool_identify_focus_areas,
        "description": "Identifies specific aspects that need further investigation.",
        "parameters": {
            "prompt": {"type": "string", "description": "The original research prompt."},
            "insights": {"type": "array", "description": "Existing insights to build upon."},
            "failed_areas": {"type": "array", "description": "Previously tried areas that yielded poor results."}
        },
    },
    "extract_key_entities": {
        "function": tool_extract_key_entities,
        "description": "Extracts key entities from the prompt for focused research.",
        "parameters": {
            "prompt": {"type": "string", "description": "The original research prompt."}
        },
    },
    "meta_analyze": {
        "function": tool_meta_analyze,
        "description": "Performs meta-analysis across entity-specific insights.",
        "parameters": {
            "entity_insights": {"type": "object", "description": "Dictionary mapping entities to their insights."},
            "prompt": {"type": "string", "description": "The original research prompt."}
        },
    },
    "draft_research_plan": {
        "function": tool_draft_research_plan,
        "description": "Creates a detailed research plan.",
        "parameters": {
            "prompt": {"type": "string", "description": "The research question/prompt."},
            "entities": {"type": "array", "description": "Key entities to investigate."},
            "focus_areas": {"type": "array", "description": "Additional areas to focus on."}
        }
    }
}

def create_prompt(task_description, user_input, available_tools, context):
    prompt = f"""{task_description}
User Input:
{user_input}
Available Tools:
"""
    for tool_name, tool_data in available_tools.items():
        prompt += f"- {tool_name}: {tool_data['description']}\n"
        prompt += "  Parameters:\n"
        for param_name, param_data in tool_data["parameters"].items():
            prompt += f"    - {param_name} ({param_data['type']}): {param_data['description']}\n"

    recent_context = context[-MAX_CONTEXT_ITEMS:] if len(context) > MAX_CONTEXT_ITEMS else context

    prompt += "\nContext (most recent items):\n"
    for item in recent_context:
        prompt += f"- {item}\n"

    prompt += """
Instructions:
Select the BEST tool and parameters for the current research stage. Output valid JSON. If no tool is appropriate, respond with {}.
Only use provided tools. Be strategic about which tool to use next based on the research progress so far.
You MUST be methodical.  Think step-by-step:
1.  **Plan:** If it's the very beginning, extract key entities, identify focus areas, and then draft a research plan.
2.  **Search:**  Use a variety of search tools.  Start with broad searches, then narrow down.  Use specific search tools (arXiv, PubMed, Scholar) for relevant topics.
3.  **Analyze:**  Reason deeply about search results, and critique your reasoning.  Identify contradictions. Filter and use FAISS index for relevant information.
4.  **Refine:** If results are poor, generate *better* search queries. Adjust focus areas.
5.  **Iterate:** Repeat steps 2-4, focusing on different entities and aspects.
6.  **Synthesize:**  Finally, summarize the findings, addressing contradictions.
Example:
{"tool": "search_web", "parameters": {"query": "Eiffel Tower location"}}
Output:
"""
    return prompt

def deep_research(prompt):
    task_description = "You are an advanced research assistant. Use available tools iteratively, focus on different aspects, follow promising leads, critically evaluate your findings, and build up a comprehensive understanding. Utilize the FAISS index to avoid redundant searches and build a persistent knowledge base."
    research_data = load_research_data()
    paper_summaries = load_paper_summaries()  # Load paper summaries

    context = research_data.get('context', [])
    all_insights = research_data.get('all_insights', [])
    entity_specific_insights = research_data.get('entity_specific_insights', {})
    intermediate_output = ""
    previous_queries = research_data.get('previous_queries', [])
    failed_queries = research_data.get('failed_queries', [])
    reasoning_context = research_data.get('reasoning_context', [])
    previous_critiques = research_data.get('previous_critiques', [])
    focus_areas = research_data.get('focus_areas', [])
    failed_areas = research_data.get('failed_areas', [])
    seen_snippets = set(research_data.get('seen_snippets', []))
    contradictions = research_data.get('contradictions', [])
    research_session_id = research_data.get('research_session_id', str(uuid4()))
    
    global index
    if research_data:
         logger.info("Restoring FAISS Index from loaded data.")
    else:
        index.reset()
        logger.info("Initialized a fresh FAISS Index")

    key_entities_with_descriptions = tool_extract_key_entities(prompt=prompt)
    key_entities = [e.split(":")[0].strip() for e in key_entities_with_descriptions] # Extract just entity names
    if key_entities:
        context.append(f"Identified key entities: {key_entities}")
        intermediate_output += f"Identified key entities for focused research: {key_entities_with_descriptions}\n"

    # Initialize progress tracking for each entity.
    entity_progress = {entity: {'queries': [], 'insights': []} for entity in key_entities}
    entity_progress['general'] = {'queries': [], 'insights': []}  # For general, non-entity-specific searches
    for entity in key_entities + ['general']:
        if entity in research_data:  # Load existing progress
            entity_progress[entity]['queries'] = research_data[entity]['queries']
            entity_progress[entity]['insights'] = research_data[entity]['insights']

    if not focus_areas:  # Corrected placement: outside the loop
        initial_focus_areas = tool_identify_focus_areas(prompt=prompt)
        research_plan = tool_draft_research_plan(prompt=prompt, entities=key_entities, focus_areas=initial_focus_areas)
        context.append(f"Initial Research Plan: {research_plan[:200]}...") # Add plan to context
        intermediate_output += f"Initial Research Plan:\n{research_plan}\n\n"
        focus_areas = initial_focus_areas


    for i in range(MAX_ITERATIONS):
        # Entity-focused iteration strategy
        if key_entities and i > 0:  # Cycle through entities *after* initial setup
            entities_to_process = key_entities + ['general']  # Include 'general' for broad searches
            current_entity = entities_to_process[i % len(entities_to_process)]
        else:
            current_entity = 'general'  # Start with general research.

        context.append(f"Current focus: {current_entity}")
        
        # FAISS Retrieval
        if i > 0:  # Use FAISS *after* the first iteration (once we have data)
            faiss_results_indices = search_faiss_index(prompt if current_entity == 'general' else f"{prompt} {current_entity}")
            faiss_context = []
            for idx in faiss_results_indices:
                if idx < len(all_insights):  # Check index bounds
                    faiss_context.append(f"Previously found insight: {all_insights[idx]}")
            if faiss_context:
                context.extend(faiss_context) # Add FAISS context
                intermediate_output += f"Iteration {i+1} - Retrieved {len(faiss_context)} relevant items from FAISS index.\n"
                

        if i == 0: #Initial broad search
            initial_query = tool_generate_search_query(prompt=prompt)
            if initial_query:
                previous_queries.append(initial_query)
                entity_progress['general']['queries'].append(initial_query)

                with ThreadPoolExecutor(max_workers=5) as executor:
                    futures = [
                        executor.submit(tool_search_web, query=initial_query, num_results=NUM_RESULTS),
                        executor.submit(tool_search_arxiv, query=initial_query, max_results=5),
                        executor.submit(tool_search_pubmed, query=initial_query, max_results=5),
                        executor.submit(tool_search_wikipedia, query=initial_query, max_results=3),
                        executor.submit(tool_search_scholar, query=initial_query, max_results=5)
                    ]

                    search_results = []
                    for future in as_completed(futures):
                        search_results.extend(future.result())

                filtered_search_results = filter_results(search_results, prompt)

                if filtered_search_results:
                    context.append(f"Initial Search Results: {len(filtered_search_results)} items found")
                    reasoning_output = tool_reason(prompt, filtered_search_results)
                    if reasoning_output:
                        all_insights.append(reasoning_output)
                        entity_progress['general']['insights'].append(reasoning_output)
                        reasoning_context.append(reasoning_output)
                        context.append(f"Initial Reasoning: {reasoning_output[:200]}...")
                        add_to_faiss_index(reasoning_output)
                else:
                    failed_queries.append(initial_query)
                    context.append(f"Initial query yielded no relevant results: {initial_query}")

        elif current_entity != 'general':
            entity_query = tool_generate_search_query(
                prompt=f"{prompt} focusing specifically on {current_entity}",
                previous_queries=entity_progress[current_entity]['queries'],
                focus_areas=focus_areas
            )

            if entity_query:
                previous_queries.append(entity_query)
                entity_progress[current_entity]['queries'].append(entity_query)


                with ThreadPoolExecutor(max_workers=5) as executor:
                    futures = [
                        executor.submit(tool_search_web, query=entity_query, num_results=NUM_RESULTS//2),
                        executor.submit(tool_search_arxiv, query=entity_query, max_results=3),
                        executor.submit(tool_search_pubmed, query=entity_query, max_results=3),
                        executor.submit(tool_search_wikipedia, query=entity_query, max_results=2),
                        executor.submit(tool_search_scholar, query=entity_query, max_results=3)
                    ]

                    search_results = []
                    for future in as_completed(futures):
                        search_results.extend(future.result())


                filtered_search_results = filter_results(search_results,
                                                        f"{prompt} {current_entity}",
                                                        previous_snippets=seen_snippets) # Pass existing snippets

                if filtered_search_results:
                    context.append(f"Entity Search for {current_entity}: {len(filtered_search_results)} results")

                    entity_reasoning = tool_reason(
                        prompt=f"{prompt} focusing on {current_entity}",
                        search_results=filtered_search_results,
                        reasoning_context=entity_progress[current_entity]['insights'], # Use entity-specific context
                        focus_areas=focus_areas
                    )

                    if entity_reasoning:
                        all_insights.append(entity_reasoning)
                        entity_progress[current_entity]['insights'].append(entity_reasoning)

                        if current_entity not in entity_specific_insights:
                            entity_specific_insights[current_entity] = []
                        entity_specific_insights[current_entity].append(entity_reasoning)

                        context.append(f"Reasoning about {current_entity}: {entity_reasoning[:200]}...")
                        add_to_faiss_index(entity_reasoning)
                else:
                    failed_queries.append(entity_query)
                    context.append(f"Entity query for {current_entity} yielded no relevant results")

        llm_prompt = create_prompt(task_description, prompt, tools, context)
        llm_response = hf_inference(MAIN_LLM_MODEL, llm_prompt)

        if isinstance(llm_response, dict) and "error" in llm_response:
            intermediate_output += f"LLM Error: {llm_response['error']}\n"
            continue

        if not isinstance(llm_response, dict) or "generated_text" not in llm_response:
            intermediate_output += "Error: Invalid LLM response.\n"
            continue

        try:
            response_text = llm_response["generated_text"].strip()
            response_json = json.loads(response_text)  # Parse the JSON response.
            intermediate_output += f"Iteration {i+1} - Focus: {current_entity} - Action: {response_text}\n"
        except json.JSONDecodeError:
            intermediate_output += f"Iteration {i+1} - LLM Response (Invalid JSON): {llm_response['generated_text'][:100]}...\n"
            context.append(f"Invalid JSON: {llm_response['generated_text'][:100]}...") # Add invalid JSON to context
            continue

        tool_name = response_json.get("tool")
        parameters = response_json.get("parameters", {})

        if not tool_name: #LLM didn't return a tool.  End the process if we are past halfway.
            if all_insights:
                if i > MAX_ITERATIONS // 2:
                    break
            continue

        if tool_name not in tools:
            context.append(f"Invalid tool: {tool_name}")
            intermediate_output += f"Iteration {i + 1} - Invalid tool chosen: {tool_name}\n"
            continue

        tool = tools[tool_name]
        try:
            intermediate_output += f"Iteration {i+1} - Executing: {tool_name}, Key params: {str(parameters)[:100]}...\n"

            if tool_name == "generate_search_query":
                parameters['previous_queries'] = previous_queries
                parameters['failed_queries'] = failed_queries
                parameters['focus_areas'] = focus_areas
                result = tool["function"](**parameters)

                if current_entity != 'general':
                    entity_progress[current_entity]['queries'].append(result) # Add entity-specific

                previous_queries.append(result)

            elif tool_name in ["search_web", "search_arxiv", "search_pubmed", "search_wikipedia", "search_scholar"]:
                result = tool["function"](**parameters)
                search_prompt = prompt
                if current_entity != 'general':
                    search_prompt = f"{prompt} focusing on {current_entity}"

                filtered_result = filter_results(result, search_prompt, previous_snippets=seen_snippets)

                result = filtered_result  # Work with filtered results

                if not result and 'query' in parameters: # Add query to failures if nothing returned.
                    failed_queries.append(parameters['query'])

            elif tool_name == "reason":
                # Ensure correct reasoning context is passed.
                if current_entity != 'general' and 'reasoning_context' not in parameters:
                    parameters['reasoning_context'] = entity_progress[current_entity]['insights']
                elif 'reasoning_context' not in parameters:
                    parameters['reasoning_context'] = reasoning_context[:]

                if 'prompt' not in parameters:
                    if current_entity != 'general':
                        parameters['prompt'] = f"{prompt} focusing on {current_entity}"
                    else:
                        parameters['prompt'] = prompt

                if 'search_results' not in parameters:
                    parameters['search_results'] = [] #Avoid errors if no search results.

                if 'focus_areas' not in parameters and focus_areas: # Avoid overwriting focus_areas if already set
                    parameters['focus_areas'] = focus_areas

                result = tool["function"](**parameters)

                if current_entity != 'general':
                    entity_progress[current_entity]['insights'].append(result)
                    if current_entity not in entity_specific_insights:
                         entity_specific_insights[current_entity] = []
                    entity_specific_insights[current_entity].append(result)
                else:
                    reasoning_context.append(result) #Add to general context.
                add_to_faiss_index(result)
                all_insights.append(result)

            elif tool_name == "critique_reasoning":
                if 'previous_critiques' not in parameters: #Pass in the previous critiques.
                    parameters['previous_critiques'] = previous_critiques

                if all_insights:
                    if 'reasoning_output' not in parameters:
                        parameters['reasoning_output'] = all_insights[-1]  #Critique the most recent insight.
                    if 'prompt' not in parameters:
                        parameters['prompt'] = prompt

                    result = tool["function"](**parameters)
                    previous_critiques.append(result)
                    context.append(f"Critique: {result[:200]}...")
                else:
                    result = "No reasoning to critique yet."

            elif tool_name == "identify_contradictions":
                result = tool["function"](**parameters)
                if result:
                    contradictions = result  # Keep track of contradictions.
                    context.append(f"Identified contradictions: {result}")

            elif tool_name == "identify_focus_areas":
                if 'failed_areas' not in parameters:
                    parameters['failed_areas'] = failed_areas
                result = tool["function"](**parameters)
                if result:
                    old_focus = set(focus_areas)
                    focus_areas = result  # Update focus areas
                    failed_areas.extend([area for area in old_focus if area not in result])  #Track failed areas
                    context.append(f"New focus areas: {result}")

            elif tool_name == "extract_article":
                result = tool["function"](**parameters)
                if result:
                    context.append(f"Extracted article content from {parameters['url']}: {result[:200]}...")
                    # Reason specifically about the extracted article.
                    reasoning_about_article = tool_reason(prompt=prompt, search_results=[{"title": "Extracted Article", "snippet": result, "url": parameters['url']}])
                    if reasoning_about_article:
                        all_insights.append(reasoning_about_article)
                        add_to_faiss_index(reasoning_about_article)
            
            elif tool_name == "summarize_paper":
                result = tool["function"](**parameters)
                if result:
                    paper_summaries[parameters['paper_text'][:100]] = result  # Store by a snippet of the text
                    save_paper_summaries(paper_summaries)
                    context.append(f"Summarized paper: {result[:200]}...")
                    add_to_faiss_index(result) # Add the summary itself to FAISS.
                    all_insights.append(result) #Add summary to insights for later summarization.

            elif tool_name == "meta_analyze":
                if 'entity_insights' not in parameters:
                    parameters['entity_insights'] = entity_specific_insights
                if 'prompt' not in parameters:
                    parameters['prompt'] = prompt
                result = tool["function"](**parameters)
                if result:
                    all_insights.append(result)  # Add meta-analysis to overall insights.
                    context.append(f"Meta-analysis across entities: {result[:200]}...")
                    add_to_faiss_index(result)


            elif tool_name == "draft_research_plan":
                result = "Research plan already generated."  # Avoid re-generating.

            else:
                result = tool["function"](**parameters)

            result_str = str(result)
            if len(result_str) > 500:
                result_str = result_str[:500] + "..."

            intermediate_output += f"Iteration {i+1} - Result: {result_str}\n"

            # Add tool use to context, limit context length
            result_context = result_str
            if len(result_str) > 300:
                result_context = result_str[:300] + "..."
            context.append(f"Used: {tool_name}, Result: {result_context}")

        except Exception as e:
            logger.error(f"Error with {tool_name}: {str(e)}")
            context.append(f"Error with {tool_name}: {str(e)}")
            intermediate_output += f"Iteration {i+1} - Error: {str(e)}\n"
            continue

        #Save data
        research_data = {
            'context': context,
            'all_insights': all_insights,
            'entity_specific_insights': entity_specific_insights,
            'previous_queries': previous_queries,
            'failed_queries': failed_queries,
            'reasoning_context': reasoning_context,
            'previous_critiques': previous_critiques,
            'focus_areas': focus_areas,
            'failed_areas': failed_areas,
            'seen_snippets': list(seen_snippets),
            'contradictions': contradictions,
            'research_session_id': research_session_id
        }
        for entity in entity_progress:
             research_data[entity] = entity_progress[entity] #save the individual entity
        save_research_data(research_data, index)


    # Perform meta-analysis *before* final summarization, if we have enough entity-specific insights.
    if len(entity_specific_insights) > 1 and len(all_insights) > 2:
        meta_analysis = tool_meta_analyze(entity_insights=entity_specific_insights, prompt=prompt)
        if meta_analysis:
            all_insights.append(meta_analysis)
            intermediate_output += f"Final Meta-Analysis: {meta_analysis[:500]}...\n"
            add_to_faiss_index(meta_analysis)  # Add to FAISS

    if all_insights:
        final_result = tool_summarize(all_insights, prompt, contradictions) # Summarize all insights.
    else:
        final_result = "Could not find meaningful information despite multiple attempts."


    full_output = f"**Research Prompt:** {prompt}\n\n"

    if key_entities_with_descriptions:
         full_output += f"**Key Entities Identified:**\n"
         for entity in key_entities_with_descriptions:
             full_output += f"- {entity}\n"
         full_output += "\n"

    full_output += "**Research Process:**\n" + intermediate_output + "\n"

    if contradictions:
        full_output += "**Contradictions Identified:**\n"
        for i, contradiction in enumerate(contradictions, 1):
            full_output += f"{i}. {contradiction}\n"
        full_output += "\n"

    full_output += f"**Final Analysis:**\n{final_result}\n\n"

    full_output += f"Research Session ID: {research_session_id}\n"
    full_output += f"Completed at: {datetime.now().strftime('%Y-%m-%d %H:%M:%S')}\n"
    full_output += f"Total iterations: {i+1}\n"
    full_output += f"Total insights generated: {len(all_insights)}\n"

    return full_output

custom_css = """
.gradio-container {
    background-color: #f7f9fc;
}
.output-box {
    font-family: 'Segoe UI', Tahoma, Geneva, Verdana, sans-serif;
    line-height: 1.5;
    font-size: 14px;
}
h3 {
    color: #2c3e50;
    font-weight: 600;
}
.footer {
    text-align: center;
    margin-top: 20px;
    color: #7f8c8d;
    font-size: 0.9em;
}
"""

iface = gr.Interface(
    fn=deep_research,
    inputs=[
        gr.Textbox(lines=5, placeholder="Enter your research question...", label="Research Question")
    ],
    outputs=gr.Textbox(lines=30, placeholder="Research results will appear here...", label="Research Results", elem_classes=["output-box"]),
    title="Advanced Multi-Stage Research Assistant",
    description="""This tool performs deep, multi-faceted research, leveraging multiple search engines, 
                   specialized academic databases, and advanced AI models. It incorporates a persistent knowledge 
                   base using FAISS indexing to avoid redundant searches and build upon previous findings.""",
    examples=[
        ["What are the key factors affecting urban tree survival and how do they vary between developing and developed countries?"],
        ["Compare and contrast the economic policies of China and the United States over the past two decades, analyzing their impacts on global trade."],
        ["What are the most promising approaches to quantum computing and what are their respective advantages and limitations?"],
        ["Analyze the environmental and social impacts of lithium mining for electric vehicle batteries."],
        ["How has artificial intelligence influenced medical diagnostics in the past five years, and what are the ethical considerations?"]
    ],
    theme="default",
    cache_examples=False,
    css=custom_css,
    allow_flagging="never",
)

if __name__ == "__main__":
    iface.launch(share=False)