import logging import os import secrets import tempfile import traceback from contextlib import asynccontextmanager from pathlib import Path from typing import Annotated, Any, Dict, List, Literal, Optional from fastapi import Depends, FastAPI, Form, HTTPException, Request, Security, UploadFile from fastapi.responses import JSONResponse from fastapi.security import APIKeyHeader from llamore import ( GeminiExtractor, LineByLinePrompter, OpenaiExtractor, References, SchemaPrompter, ) from pydantic import BaseModel, BeforeValidator, Field logging.basicConfig( level=logging.INFO, format="%(asctime)s - %(name)s - %(levelname)s - %(message)s", handlers=[logging.StreamHandler()], ) logger = logging.getLogger(__name__) # ===== Config ===== ALLOWED_API_KEY = os.getenv("ALLOWED_API_KEY") if not ALLOWED_API_KEY: raise ValueError("ALLOWED_API_KEY environment variable must be set") MAX_PDF_SIZE_BYTES = int(os.getenv("MAX_PDF_SIZE_MB", "50")) * 1024 * 1024 # ===== Types ===== def _coerce_dict(v: Any) -> Optional[Dict[str, Any]]: """Accept a dict, None, or empty string; reject anything else.""" if v is None or v == "": return None if isinstance(v, dict): return v raise ValueError(f"Expected a JSON object, got {type(v).__name__!r}") OptionalDict = Annotated[Optional[Dict[str, Any]], BeforeValidator(_coerce_dict)] # ===== Auth ===== api_key_header = APIKeyHeader(name="X-Llamore-API-Key", scheme_name="Llamore API Key", auto_error=False) provider_key_header = APIKeyHeader(name="X-LLM-Provider-Key", scheme_name="LLM Provider Key", auto_error=False) def api_error(detail: str, status_code: int = 400) -> HTTPException: """Create an HTTPException with server-side logging.""" logger.error(detail) return HTTPException(status_code=status_code, detail=detail) async def verify_api_key(api_key: str = Security(api_key_header)): if not api_key or not secrets.compare_digest(api_key, ALLOWED_API_KEY): raise HTTPException(status_code=401, detail="Invalid or missing API key") return api_key async def verify_provider_key(provider_api_key: str = Security(provider_key_header)): if not provider_api_key or not provider_api_key.strip(): raise HTTPException(status_code=401, detail="Missing or empty provider API key") return provider_api_key # ===== App ===== @asynccontextmanager async def lifespan(app: FastAPI): logger.info("Starting llamore FastAPI application") yield logger.info("Shutting down llamore FastAPI application") app = FastAPI( title="Llamore API", description="API for extracting and processing scholarly references using llamore", version="1.0.0", lifespan=lifespan, ) @app.exception_handler(Exception) async def global_exception_handler(request: Request, exc: Exception): if isinstance(exc, HTTPException): raise exc logger.error( f"Unhandled exception in {request.method} {request.url.path}:\n{traceback.format_exc()}" ) return JSONResponse( status_code=500, content={"detail": "An internal server error occurred."}, ) # ===== Schemas ===== class BaseExtractionConfig(BaseModel): """Options shared across all providers and input types.""" prompter_type: Literal["schema", "line_by_line"] = Field( "schema", description="Prompter type for extraction.", ) step_by_step: bool = Field( False, description="Enable step-by-step extraction (SchemaPrompter only).", ) extra_api_kwargs: OptionalDict = Field( None, description="Extra keyword arguments forwarded to the provider's generate call.", ) return_xml: bool = Field( False, description="If true, also return a TEI XML representation of the extracted references.", ) class OpenaiExtractionConfig(BaseExtractionConfig): """OpenAI-specific extraction options.""" model: str = Field("gpt-4o", description="OpenAI model name.") endpoint: Literal["create", "parse"] = Field( "create", description=( "'parse' uses beta.chat.completions.parse for native structured output " "and requires a compatible model. " "Cannot be combined with prompter_type='line_by_line'." ), ) client_kwargs: OptionalDict = Field( None, description=( "Extra keyword arguments forwarded to the openai.OpenAI() constructor " "(e.g. base_url for Ollama/vLLM/SGLang-compatible endpoints, " "timeout, max_retries, default_headers)." ), ) @classmethod def as_form( cls, model: str = Form("gpt-4o"), prompter_type: Literal["schema", "line_by_line"] = Form("schema"), step_by_step: bool = Form(False), endpoint: Literal["create", "parse"] = Form("create"), client_kwargs: OptionalDict = Form(None), extra_api_kwargs: OptionalDict = Form(None), return_xml: bool = Form(False), ) -> "OpenaiExtractionConfig": return cls( model=model, prompter_type=prompter_type, step_by_step=step_by_step, endpoint=endpoint, client_kwargs=client_kwargs, extra_api_kwargs=extra_api_kwargs, return_xml=return_xml, ) class GeminiExtractionConfig(BaseExtractionConfig): """Gemini-specific extraction options.""" model: str = Field("gemini-2.5-flash", description="Gemini model name.") @classmethod def as_form( cls, model: str = Form("gemini-2.5-flash"), prompter_type: Literal["schema", "line_by_line"] = Form("schema"), step_by_step: bool = Form(False), extra_api_kwargs: OptionalDict = Form(None), return_xml: bool = Form(False), ) -> "GeminiExtractionConfig": return cls( model=model, prompter_type=prompter_type, step_by_step=step_by_step, extra_api_kwargs=extra_api_kwargs, return_xml=return_xml, ) class OpenaiExtractTextRequest(OpenaiExtractionConfig): """Request body for OpenAI text extraction.""" text: str = Field(..., min_length=1, description="Raw text to extract references from.") class GeminiExtractTextRequest(GeminiExtractionConfig): """Request body for Gemini text extraction.""" text: str = Field(..., min_length=1, description="Raw text to extract references from.") class ReferencesResponse(BaseModel): """Response containing extracted references and optional TEI XML.""" references: List[Dict[str, Any]] = Field( ..., description="List of extracted references as JSON objects.", ) xml: Optional[str] = Field( None, description="TEI XML representation of references (only present if return_xml=True).", ) # ===== Factories ===== def _build_prompter( prompter_type: Literal["schema", "line_by_line"], step_by_step: bool, endpoint: Literal["create", "parse"] = "create", ): if prompter_type == "line_by_line": if endpoint == "parse": raise api_error( "The 'parse' endpoint is incompatible with the 'line_by_line' prompter." ) return LineByLinePrompter() elif prompter_type == "schema": return SchemaPrompter(step_by_step=step_by_step) else: raise api_error( f"Unsupported prompter_type '{prompter_type}'. Choose 'schema' or 'line_by_line'." ) def create_openai_extractor( provider_api_key: str, config: OpenaiExtractionConfig, ) -> OpenaiExtractor: prompter = _build_prompter(config.prompter_type, config.step_by_step, config.endpoint) return OpenaiExtractor( api_key=provider_api_key, model=config.model, prompter=prompter, endpoint=config.endpoint, **(config.client_kwargs or {}), ) def create_gemini_extractor( provider_api_key: str, config: GeminiExtractionConfig, ) -> GeminiExtractor: prompter = _build_prompter(config.prompter_type, config.step_by_step) return GeminiExtractor( api_key=provider_api_key, model=config.model, prompter=prompter, ) def references_to_response(references: References, return_xml: bool) -> ReferencesResponse: refs_dict = [ref.model_dump(exclude_none=True) for ref in references] xml: Optional[str] = None if return_xml and references: try: xml = references.to_xml(pretty_print=True) except Exception: logger.warning("Failed to convert references to TEI XML.", exc_info=True) return ReferencesResponse(references=refs_dict, xml=xml) async def _read_and_validate_pdf(file: UploadFile) -> bytes: if not file.filename or not file.filename.lower().endswith(".pdf"): raise api_error("A valid .pdf file is required.") content = await file.read() if not content: raise api_error("Uploaded file is empty.") if len(content) > MAX_PDF_SIZE_BYTES: raise api_error( f"PDF exceeds the maximum allowed size of {MAX_PDF_SIZE_BYTES // (1024 * 1024)} MB.", status_code=413, ) return content async def _run_pdf_extraction( extractor, file: UploadFile, extra_api_kwargs: OptionalDict, return_xml: bool, ) -> ReferencesResponse: content = await _read_and_validate_pdf(file) tmp_path: Optional[Path] = None try: with tempfile.NamedTemporaryFile(delete=False, suffix=".pdf") as tmp: tmp.write(content) tmp_path = Path(tmp.name) try: references = extractor(pdf=tmp_path, **(extra_api_kwargs or {})) except HTTPException: raise except Exception: logger.error("PDF extraction failed for '%s'.", file.filename, exc_info=True) raise api_error("Reference extraction failed. Check server logs for details.") finally: if tmp_path and tmp_path.exists(): try: tmp_path.unlink() except Exception: logger.warning("Could not delete temporary file '%s'.", tmp_path, exc_info=True) logger.info("Extracted %d references from '%s'.", len(references), file.filename) return references_to_response(references, return_xml) # ===== Endpoints ===== @app.get("/") async def root(): return { "message": "Llamore API", "version": "1.0.0", "endpoints": { "extract_openai_text": "/extract/openai/text", "extract_openai_pdf": "/extract/openai/pdf", "extract_gemini_text": "/extract/gemini/text", "extract_gemini_pdf": "/extract/gemini/pdf", "health": "/health", }, } @app.get("/health") async def health_check(): return {"status": "healthy", "service": "llamore-api"} @app.post("/extract/openai/text", response_model=ReferencesResponse) async def extract_openai_text( request: OpenaiExtractTextRequest, provider_api_key: str = Security(verify_provider_key), _: str = Security(verify_api_key), ): """Extract references from plain text using OpenAI.""" if not request.text.strip(): raise api_error("Text cannot be empty.") try: extractor = create_openai_extractor(provider_api_key, request) references = extractor( text=request.text, **(request.extra_api_kwargs or {}), ) except HTTPException: raise except Exception: logger.error("Text extraction failed.", exc_info=True) raise api_error("Reference extraction failed. Check server logs for details.") logger.info("Extracted %d references from text.", len(references)) return references_to_response(references, request.return_xml) @app.post("/extract/openai/pdf", response_model=ReferencesResponse) async def extract_openai_pdf( file: UploadFile, config: OpenaiExtractionConfig = Depends(OpenaiExtractionConfig.as_form), provider_api_key: str = Security(verify_provider_key), _: str = Security(verify_api_key), ): """Extract references from a PDF file using OpenAI.""" try: extractor = create_openai_extractor(provider_api_key, config) except HTTPException: raise return await _run_pdf_extraction(extractor, file, config.extra_api_kwargs, config.return_xml) @app.post("/extract/gemini/text", response_model=ReferencesResponse) async def extract_gemini_text( request: GeminiExtractTextRequest, provider_api_key: str = Security(verify_provider_key), _: str = Security(verify_api_key), ): """Extract references from plain text using Gemini.""" if not request.text.strip(): raise api_error("Text cannot be empty.") try: extractor = create_gemini_extractor(provider_api_key, request) references = extractor( text=request.text, **(request.extra_api_kwargs or {}), ) except HTTPException: raise except Exception: logger.error("Text extraction failed.", exc_info=True) raise api_error("Reference extraction failed. Check server logs for details.") logger.info("Extracted %d references from text.", len(references)) return references_to_response(references, request.return_xml) @app.post("/extract/gemini/pdf", response_model=ReferencesResponse) async def extract_gemini_pdf( file: UploadFile, config: GeminiExtractionConfig = Depends(GeminiExtractionConfig.as_form), provider_api_key: str = Security(verify_provider_key), _: str = Security(verify_api_key), ): """Extract references from a PDF file using Gemini.""" try: extractor = create_gemini_extractor(provider_api_key, config) except HTTPException: raise return await _run_pdf_extraction(extractor, file, config.extra_api_kwargs, config.return_xml)