mcp/orchestrator.py

import asyncio
import httpx
from typing import Any, Dict, List, Literal, Union

from mcp.pubmed import fetch_pubmed
from mcp.arxiv import fetch_arxiv
from mcp.umls import extract_umls_concepts
from mcp.openfda import fetch_drug_safety
from mcp.ncbi import search_gene, get_mesh_definition
from mcp.mygene import fetch_gene_info
from mcp.ensembl import fetch_ensembl
from mcp.opentargets import fetch_ot
from mcp.clinicaltrials import search_trials
from mcp.cbio import fetch_cbio
from mcp.disgenet import disease_to_genes
from mcp.gemini import gemini_summarize, gemini_qa
from mcp.openai_utils import ai_summarize, ai_qa


async def _safe_call(
    func: Any,
    *args,
    default: Any = None,
    **kwargs,
) -> Any:
    """
    Safely call an async function, returning a default on HTTP or other failures.
    """
    try:
        return await func(*args, **kwargs)  # type: ignore
    except httpx.HTTPStatusError:
        return default
    except Exception:
        return default


async def _gather_tasks(tasks: List[asyncio.Task]) -> List[Any]:
    """
    Await a list of asyncio.Tasks and return their results in order.
    """
    return await asyncio.gather(*tasks)


def _flatten_unique(items: List[Union[List[Any], Any]]) -> List[Any]:
    """
    Flatten a list of items where elements may be lists or single values,
    then deduplicate preserving order.
    """
    flat: List[Any] = []
    for elem in items:
        if isinstance(elem, list):
            for x in elem:
                if x not in flat:
                    flat.append(x)
        else:
            if elem is not None and elem not in flat:
                flat.append(elem)
    return flat


async def orchestrate_search(
    query: str,
    llm: Literal['openai', 'gemini'] = 'openai',
    max_papers: int = 7,
    max_trials: int = 10,
) -> Dict[str, Any]:
    """
    Perform a comprehensive biomedical search pipeline with fault tolerance:
      - Literature (PubMed + arXiv)
      - Entity extraction (UMLS)
      - Drug safety, gene & variant info, disease-gene mapping
      - Clinical trials, cBioPortal data
      - AI-driven summary

    Individual fetch functions that fail with an HTTP error will return an empty default,
    ensuring the pipeline always completes.
    """
    tasks = {
        'pubmed': asyncio.create_task(fetch_pubmed(query, max_results=max_papers)),
        'arxiv': asyncio.create_task(fetch_arxiv(query, max_results=max_papers)),
        'umls': asyncio.create_task(
            asyncio.to_thread(extract_umls_concepts, query)
        ),
        'drug_safety': asyncio.create_task(_safe_call(fetch_drug_safety, query, default=[])),
        'ncbi_gene': asyncio.create_task(_safe_call(search_gene, query, default=[])),
        'mygene': asyncio.create_task(_safe_call(fetch_gene_info, query, default=[])),
        'ensembl': asyncio.create_task(_safe_call(fetch_ensembl, query, default=[])),
        'opentargets': asyncio.create_task(_safe_call(fetch_ot, query, default=[])),
        'mesh': asyncio.create_task(_safe_call(get_mesh_definition, query, default="")),
        'trials': asyncio.create_task(_safe_call(search_trials, query, default=[], max_studies=max_trials)),
        'cbio': asyncio.create_task(_safe_call(fetch_cbio, query, default=[])),
        'disgenet': asyncio.create_task(_safe_call(disease_to_genes, query, default=[])),
    }

    results = await _gather_tasks(list(tasks.values()))
    data = dict(zip(tasks.keys(), results))

    gene_sources = [data['ncbi_gene'], data['mygene'], data['ensembl'], data['opentargets']]
    genes = _flatten_unique(gene_sources)

    papers = (data['pubmed'] or []) + (data['arxiv'] or [])

    summaries = " ".join(p.get('summary', '') for p in papers)
    if llm == 'gemini':
        ai_summary = await gemini_summarize(summaries)
        llm_used = 'gemini'
    else:
        ai_summary = await ai_summarize(summaries)
        llm_used = 'openai'

    return {
        'papers': papers,
        'genes': genes,
        'umls': data['umls'] or [],
        'gene_disease': data['disgenet'] or [],
        'mesh_defs': [data['mesh']] if data['mesh'] else [],
        'drug_safety': data['drug_safety'] or [],
        'clinical_trials': data['trials'] or [],
        'variants': data['cbio'] or [],
        'ai_summary': ai_summary,
        'llm_used': llm_used,
    }


async def answer_ai_question(
    question: str,
    context: str = "",
    llm: Literal['openai', 'gemini'] = 'openai',
) -> Dict[str, str]:
    """
    Answer a free-text question using the specified LLM, with fallback.
    Returns {'answer': text}.
    """
    try:
        if llm == 'gemini':
            answer = await gemini_qa(question, context)
        else:
            answer = await ai_qa(question, context)
    except Exception as e:
        answer = f"LLM unavailable or quota exceeded: {e}"
    return {'answer': answer}