Skip to content

BrainAPI is a knowledge graph–powered AI memory layer that transforms unstructured data into structured knowledge, enabling intelligent search, recommendations, and contextual memory for AI agents and applications.

2.4 KB
Raw Blame
import json

import boto3
from langchain_aws import ChatBedrockConverse

from src.adapters.interfaces.llm import LLM
from src.config import config


class _BedrockLLMBase(LLM):
    def __init__(self, model: str):
        self.model = model
        session = boto3.Session(
            aws_access_key_id=config.bedrock.access_key_id,
            aws_secret_access_key=config.bedrock.secret_access_key,
            aws_session_token=config.bedrock.session_token,
            region_name=config.bedrock.region,
        )
        self.client = session.client("bedrock-runtime")
        self.langchain_model = ChatBedrockConverse(
            model=model,
            region_name=config.bedrock.region,
            aws_access_key_id=config.bedrock.access_key_id,
            aws_secret_access_key=config.bedrock.secret_access_key,
            aws_session_token=config.bedrock.session_token,
        )

    def _converse(self, prompt: str, max_new_tokens: int = None) -> str:
        payload = {
            "modelId": self.model,
            "messages": [{"role": "user", "content": [{"text": prompt}]}],
        }
        if max_new_tokens:
            payload["inferenceConfig"] = {"maxTokens": max_new_tokens}
        response = self.client.converse(**payload)
        content = response["output"]["message"]["content"]
        texts = [part.get("text", "") for part in content if isinstance(part, dict)]
        return "".join(texts).strip()

    def generate_text(self, prompt: str, max_new_tokens: int = None) -> str:
        return self._converse(prompt, max_new_tokens=max_new_tokens)

    def generate_json(
        self, prompt: str, max_new_tokens: int = None, max_retries: int = 3
    ) -> dict:
        while max_retries > 0:
            try:
                response = self._converse(prompt, max_new_tokens=max_new_tokens)
                raw = response.strip().strip("```json").strip("```").strip()
                return json.loads(raw)
            except Exception:
                max_retries -= 1
                if max_retries <= 0:
                    raise
        return {}


class BedrockLLMClientSmall(_BedrockLLMBase):
    def __init__(self):
        super().__init__(config.bedrock.small_llm_model)


class BedrockLLMClientLarge(_BedrockLLMBase):
    def __init__(self):
        super().__init__(config.bedrock.large_llm_model)


_llm_small_client_bedrock = BedrockLLMClientSmall()
_llm_large_client_bedrock = BedrockLLMClientLarge()