"""OpenAI-compatible LLM adapter for trail guide extraction. Works with any OpenAI-compatible endpoint (Ollama, Groq, Gemini, OpenAI) via configurable base_url / model / api_key. """ from __future__ import annotations import json import logging from typing import Any from openai import OpenAI from app.config import LLMConfig log = logging.getLogger("trekker.llm") EXTRACTION_PROMPT = """You extract structured overland trail data from text. Return ONLY a JSON object with these keys (use null/[] when not mentioned): { "name": string, "difficulty": string|null, // e.g. "easy", "moderate", "difficult", "extreme", or a rating like "7/10" "obstacle_types": [string], // e.g. "rock garden", "ledges", "mud", "off-camber" "water_crossings": [{"name": string|null, "depth_ft": number|null, "notes": string|null}], "gate_codes": [string], "vehicle_recommendations": [string],// e.g. "33-inch tires", "lockers recommended", "winch" "closures": [string], // seasonal or permanent closure notes "warnings": [string], "notes": string|null }""" class LLMService: def __init__(self, config: LLMConfig): self.config = config @property def configured(self) -> bool: return bool(self.config.api_key or self.config.base_url) def _client(self) -> OpenAI: return OpenAI( base_url=self.config.base_url or None, api_key=self.config.api_key or "not-needed", # local providers ignore the key ) def extract_trail_data(self, text: str) -> dict[str, Any]: """Trail guide text -> structured dict. Raises on provider errors.""" client = self._client() # Keep prompts inside small local-model context windows. snippet = text[:24000] kwargs: dict[str, Any] = { "model": self.config.model, "messages": [ {"role": "system", "content": EXTRACTION_PROMPT}, {"role": "user", "content": snippet}, ], } try: resp = client.chat.completions.create( **kwargs, response_format={"type": "json_object"} ) except Exception: # Some providers reject response_format; retry without it. resp = client.chat.completions.create(**kwargs) content = resp.choices[0].message.content or "{}" # Tolerate code-fenced JSON from smaller models. content = content.strip() if content.startswith("```"): content = content.strip("`") content = content.split("\n", 1)[1] if "\n" in content else content start, end = content.find("{"), content.rfind("}") if start != -1 and end > start: content = content[start : end + 1] return json.loads(content)