106 lines
4.1 KiB
Python
106 lines
4.1 KiB
Python
"""
|
|
@provider GeminiProvider (swarm/providers/gemini_provider.py)
|
|
@purpose Async HTTP provider for Google Gemini models using httpx.
|
|
"""
|
|
import time
|
|
import json
|
|
from typing import Type, TypeVar, List, Dict, Tuple
|
|
import httpx
|
|
from pydantic import BaseModel
|
|
from swarm.config import swarm_settings
|
|
from swarm.providers.base import BaseLLMProvider, TokenUsage
|
|
|
|
T = TypeVar("T", bound=BaseModel)
|
|
|
|
|
|
class GeminiProvider(BaseLLMProvider):
|
|
def __init__(self, api_key: str = None, model: str = None):
|
|
self.api_key = api_key or swarm_settings.GEMINI_API_KEY
|
|
self.model = model or swarm_settings.GEMINI_MODEL
|
|
self.base_url = "https://generativelanguage.googleapis.com/v1beta"
|
|
|
|
async def generate_structured(
|
|
self,
|
|
prompt: str,
|
|
schema: Type[T],
|
|
system_prompt: str = ""
|
|
) -> Tuple[T, TokenUsage]:
|
|
start = time.perf_counter()
|
|
url = f"{self.base_url}/models/{self.model}:generateContent?key={self.api_key}"
|
|
|
|
payload = {
|
|
"contents": [{"parts": [{"text": prompt}]}],
|
|
"generationConfig": {
|
|
"responseMimeType": "application/json",
|
|
"responseSchema": schema.model_json_schema()
|
|
}
|
|
}
|
|
if system_prompt:
|
|
payload["systemInstruction"] = {"parts": [{"text": system_prompt}]}
|
|
|
|
async with httpx.AsyncClient(timeout=60.0) as client:
|
|
resp = await client.post(url, json=payload)
|
|
resp.raise_for_status()
|
|
data = resp.json()
|
|
|
|
elapsed_ms = int((time.perf_counter() - start) * 1000)
|
|
raw_text = data["candidates"][0]["content"]["parts"][0]["text"]
|
|
parsed = schema.model_validate_json(raw_text)
|
|
|
|
usage_meta = data.get("usageMetadata", {})
|
|
prompt_tokens = usage_meta.get("promptTokenCount", len(prompt.split()) * 2)
|
|
completion_tokens = usage_meta.get("candidatesTokenCount", len(raw_text.split()) * 2)
|
|
|
|
pricing = swarm_settings.PRICING_MATRIX.get(self.model, {"input_per_million": 0.10, "output_per_million": 0.40})
|
|
cost = (prompt_tokens * pricing["input_per_million"] + completion_tokens * pricing["output_per_million"]) / 1_000_000.0
|
|
|
|
usage = TokenUsage(
|
|
prompt_tokens=prompt_tokens,
|
|
completion_tokens=completion_tokens,
|
|
total_tokens=prompt_tokens + completion_tokens,
|
|
latency_ms=elapsed_ms,
|
|
model_name=self.model,
|
|
estimated_cost_usd=round(cost, 6)
|
|
)
|
|
return parsed, usage
|
|
|
|
async def chat(
|
|
self,
|
|
messages: List[Dict[str, str]],
|
|
system_prompt: str = ""
|
|
) -> Tuple[str, TokenUsage]:
|
|
start = time.perf_counter()
|
|
url = f"{self.base_url}/models/{self.model}:generateContent?key={self.api_key}"
|
|
|
|
contents = []
|
|
for msg in messages:
|
|
role = "user" if msg["role"] == "user" else "model"
|
|
contents.append({"role": role, "parts": [{"text": msg["content"]}]})
|
|
|
|
payload = {"contents": contents}
|
|
if system_prompt:
|
|
payload["systemInstruction"] = {"parts": [{"text": system_prompt}]}
|
|
|
|
async with httpx.AsyncClient(timeout=60.0) as client:
|
|
resp = await client.post(url, json=payload)
|
|
resp.raise_for_status()
|
|
data = resp.json()
|
|
|
|
elapsed_ms = int((time.perf_counter() - start) * 1000)
|
|
raw_text = data["candidates"][0]["content"]["parts"][0]["text"]
|
|
usage_meta = data.get("usageMetadata", {})
|
|
prompt_tokens = usage_meta.get("promptTokenCount", 100)
|
|
completion_tokens = usage_meta.get("candidatesTokenCount", 50)
|
|
|
|
pricing = swarm_settings.PRICING_MATRIX.get(self.model, {"input_per_million": 0.10, "output_per_million": 0.40})
|
|
cost = (prompt_tokens * pricing["input_per_million"] + completion_tokens * pricing["output_per_million"]) / 1_000_000.0
|
|
|
|
usage = TokenUsage(
|
|
prompt_tokens=prompt_tokens,
|
|
completion_tokens=completion_tokens,
|
|
total_tokens=prompt_tokens + completion_tokens,
|
|
latency_ms=elapsed_ms,
|
|
model_name=self.model,
|
|
estimated_cost_usd=round(cost, 6)
|
|
)
|
|
return raw_text, usage
|