refactor(api): add LLM abstraction layer to replace hardcoded model strings

Replace hardcoded "google_genai:gemini-3-flash-preview" model references across
all agent modules (supervisor, websearch, answer extractor, ocr reader) with
configurable get_llm() function from agent.api.api module

.env.example

@@ -1,2 +1,4 @@
 GOOGLE_API_KEY=your-google-api-key-here
 TAVILY_API_KEY=your-tavily-api-key-here
+MINIMAX_API_KEY=your-minimax-api-key-here
+MINIMAX_GROUP_ID=your-minimax-group-id-here

agent/agent.py

@@ -3,6 +3,7 @@ from dotenv import load_dotenv
 from colorama import Fore, Style  # type: ignore[import]
 from langchain.agents import create_agent
 from langchain_core.messages import HumanMessage
+from agent.api.api import get_llm
 from agent.tools.math_solver import math_solver
 from agent.tools.file_downloader import file_downloader
 from agent.tools.ocr_reader import ocr_reader
@@ -20,7 +21,7 @@ load_dotenv()
 def supervisor_agent():
     """Return a supervisor agent instance with math_solver and websearch_agent."""
     return create_agent(
-        model="google_genai:gemini-3-flash-preview",
+        model=get_llm(),
         # tools=[math_solver, websearch_agent, web_search_agents],
         tools=[math_solver, web_search_agents, file_downloader, ocr_reader, list_files, http_get],
         system_prompt=(

agent/agents/answer_extractor.py

@@ -2,6 +2,8 @@ from colorama import Fore, Style  # type: ignore[import]
 from langchain.agents import create_agent
 from pydantic import BaseModel, Field
 
+from agent.api.api import get_llm
+
 
 class ExtractedAnswer(BaseModel):
     """Structured output for answer extraction."""
@@ -19,7 +21,7 @@ def extract_answer(raw_answer: str, question: str) -> str:
     print(f"{Fore.CYAN}[AnswerExtractor] Extracting...{Style.RESET_ALL}")
 
     agent = create_agent(
-        model="google_genai:gemini-3-flash-preview",
+        model=get_llm(),
         response_format=ExtractedAnswer,
         system_prompt="""\
 You are part of a system that demands extreme precision.

agent/agents/websearchagent.py

@@ -3,6 +3,7 @@ from colorama import Fore, Style  # type: ignore[import]
 from langchain_core.tools import tool
 from langchain.agents import create_agent
 from langgraph.errors import GraphRecursionError
+from agent.api.api import get_llm
 from agent.tools.search import web_search
 
 
@@ -26,7 +27,7 @@ def websearch_agent(query: str) -> str:
     """
     print(f"{Fore.YELLOW}[SupervisorAgent -> WebSearchAgent] {query}{Style.RESET_ALL}")
     base_agent = create_agent(
-        model="google_genai:gemini-3-flash-preview",
+        model=get_llm(),
         tools=[web_search],
         system_prompt=(
             f"Current time is: {datetime.now(timezone.utc).isoformat()}. "

agent/agents/websearchagents.py

@@ -6,6 +6,8 @@ from colorama import Fore, Style  # type: ignore[import]
 from langchain.agents import create_agent
 from langchain_core.tools import tool
 from pydantic import BaseModel, Field
+
+from agent.api.api import get_llm
 from tavily import TavilyClient  # type: ignore[import]
 
 
@@ -45,7 +47,7 @@ def expand_queries(origin_question: str, query: str) -> list[str]:
     print(f"{Fore.CYAN}[QueryExpander] Expanding: {query}{Style.RESET_ALL}")
 
     agent = create_agent(
-        model="google_genai:gemini-3-flash-preview",
+        model=get_llm(),
         response_format=ExpandedQueries,
         system_prompt=(
             f"Current time is: {datetime.now(timezone.utc).isoformat()}. "
@@ -157,7 +159,7 @@ def subagent(origin_question: str, query: str) -> str:
     print(f"{color}[SubAgent] Investigating ({len(prompt)} chars)...{Style.RESET_ALL}")
 
     agent = create_agent(
-        model="google_genai:gemini-3-flash-preview",
+        model=get_llm(),
         system_prompt=(
             f"Current time is: {datetime.now(timezone.utc).isoformat()}. "
             "You are a research analyst. You are given a web page's full content "
@@ -194,7 +196,7 @@ def combine_result_agent(origin_question: str, query: str) -> str:
     )
 
     agent = create_agent(
-        model="google_genai:gemini-3-flash-preview",
+        model=get_llm(),
         system_prompt=(
             f"Current time is: {datetime.now(timezone.utc).isoformat()}. "
             "You are a research synthesizer. You receive findings from multiple "

agent/api/api.py

@@ -0,0 +1,14 @@
+# from langchain_google_genai import ChatGoogleGenerativeAI
+
+from langchain_community.chat_models import MiniMaxChat
+
+_llm = None
+
+
+def get_llm():
+    """Return a shared LLM instance."""
+    global _llm
+    if _llm is None:
+        # _llm = ChatGoogleGenerativeAI(model="gemini-3-flash-preview")
+        _llm = MiniMaxChat(model="m2.5-highspeed")
+    return _llm

agent/tools/ocr_reader.py

@@ -4,7 +4,7 @@ from pathlib import Path
 from colorama import Fore, Style  # type: ignore[import]
 from langchain_core.messages import HumanMessage
 from langchain_core.tools import tool
-from langchain_google_genai import ChatGoogleGenerativeAI
+from agent.api.api import get_llm
 
 # Workspace directory where downloaded files are stored
 WORKSPACE_DIR = Path(__file__).resolve().parents[2] / "workspace"
@@ -80,7 +80,7 @@ def ocr_reader(filename: str, query: str = "") -> str:
         base64_image = _encode_image(str(path))
         mime_type = _guess_mime_type(str(path))
 
-        llm = ChatGoogleGenerativeAI(model="gemini-3-flash-preview")
+        llm = get_llm()
 
         prompt = f"{SYSTEM_PROMPT}\n\nUser question: {query}" if query else SYSTEM_PROMPT
 

requirements.txt

@@ -5,6 +5,7 @@ python-dotenv==1.2.2
 langchain
 langchain-core
 langchain-google-genai
+langchain-community
 langgraph
 tavily-python
 sympy