Update agent flow and frontend chat experience

README.md

@@ -1,6 +1,6 @@
-# OpenAI-Compatible Fullstack LangGraph Quickstart
+# Gemini Fullstack LangGraph Quickstart
 
-This project demonstrates a fullstack application using a React frontend and a LangGraph-powered backend agent. The agent performs comprehensive research on a user's query by dynamically generating search terms, querying the web with Tavily, reflecting on the results to identify knowledge gaps, and iteratively refining its search until it can provide a well-supported answer with citations. This application serves as an example of building research-augmented conversational AI using LangGraph and an OpenAI-compatible chat model.
+This project demonstrates a fullstack application using a React frontend and a LangGraph-powered backend agent. The agent is designed to perform comprehensive research on a user's query by dynamically generating search terms, querying the web using Google Search, reflecting on the results to identify knowledge gaps, and iteratively refining its search until it can provide a well-supported answer with citations. This application serves as an example of building research-augmented conversational AI using LangGraph and Google's Gemini models.
 
 <img src="./app.png" title="Gemini Fullstack LangGraph" alt="Gemini Fullstack LangGraph" width="90%">
 
@@ -8,8 +8,8 @@ This project demonstrates a fullstack application using a React frontend and a L
 
 - 💬 Fullstack application with a React frontend and LangGraph backend.
 - 🧠 Powered by a LangGraph agent for advanced research and conversational AI.
-- 🔍 Dynamic search query generation using an OpenAI-compatible chat model.
+- 🔍 Dynamic search query generation using Google Gemini models.
-- 🌐 Integrated web research via Tavily Search.
+- 🌐 Integrated web research via Google Search API.
 - 🤔 Reflective reasoning to identify knowledge gaps and refine searches.
 - 📄 Generates answers with citations from gathered sources.
 - 🔄 Hot-reloading for both frontend and backend during development.
@@ -29,15 +29,10 @@ Follow these steps to get the application running locally for development and te
 
 -   Node.js and npm (or yarn/pnpm)
 -   Python 3.11+
--   **`OPENAI_API_KEY`**: The backend agent requires an OpenAI-compatible API key.
+-   **`GEMINI_API_KEY`**: The backend agent requires a Google Gemini API key.
--   **`OPENAI_BASE_URL`**: Set this to `https://open.bigmodel.cn/api/paas/v4` for BigModel.
--   **`TAVILY_API_KEY`**: Required for web search.
     1.  Navigate to the `backend/` directory.
     2.  Create a file named `.env` by copying the `backend/.env.example` file.
-    3.  Open the `.env` file and add:
+    3.  Open the `.env` file and add your Gemini API key: `GEMINI_API_KEY="YOUR_ACTUAL_API_KEY"`
-        `OPENAI_API_KEY="YOUR_ACTUAL_API_KEY"`
-        `OPENAI_BASE_URL="https://open.bigmodel.cn/api/paas/v4"`
-        `TAVILY_API_KEY="YOUR_ACTUAL_TAVILY_API_KEY"`
 
 **2. Install Dependencies:**
 
@@ -72,11 +67,11 @@ The core of the backend is a LangGraph agent defined in `backend/src/agent/graph
 
 <img src="./agent.png" title="Agent Flow" alt="Agent Flow" width="50%">
 
-1.  **Generate Initial Queries:** Based on your input, it generates a set of initial search queries using the configured chat model.
+1.  **Generate Initial Queries:** Based on your input, it generates a set of initial search queries using a Gemini model.
-2.  **Web Research:** For each query, it uses Tavily Search to gather relevant web pages and snippets.
+2.  **Web Research:** For each query, it uses the Gemini model with the Google Search API to find relevant web pages.
-3.  **Reflection & Knowledge Gap Analysis:** The agent analyzes the search results to determine if the information is sufficient or if there are knowledge gaps. It uses the configured chat model for this reflection process.
+3.  **Reflection & Knowledge Gap Analysis:** The agent analyzes the search results to determine if the information is sufficient or if there are knowledge gaps. It uses a Gemini model for this reflection process.
 4.  **Iterative Refinement:** If gaps are found or the information is insufficient, it generates follow-up queries and repeats the web research and reflection steps (up to a configured maximum number of loops).
-5.  **Finalize Answer:** Once the research is deemed sufficient, the agent synthesizes the gathered information into a coherent answer, including citations from the web sources, using the configured chat model.
+5.  **Finalize Answer:** Once the research is deemed sufficient, the agent synthesizes the gathered information into a coherent answer, including citations from the web sources, using a Gemini model.
 
 ## CLI Example
 
@@ -102,12 +97,12 @@ _Note: If you are not running the docker-compose.yml example or exposing the bac
 
    Run the following command from the **project root directory**:
    ```bash
-   docker build -t openai-fullstack-langgraph -f Dockerfile .
+   docker build -t gemini-fullstack-langgraph -f Dockerfile .
    ```
 **2. Run the Production Server:**
 
    ```bash
-   OPENAI_API_KEY=<your_api_key> OPENAI_BASE_URL=https://open.bigmodel.cn/api/paas/v4 TAVILY_API_KEY=<your_tavily_api_key> LANGSMITH_API_KEY=<your_langsmith_api_key> docker-compose up
+   GEMINI_API_KEY=<your_gemini_api_key> LANGSMITH_API_KEY=<your_langsmith_api_key> docker-compose up
    ```
 
 Open your browser and navigate to `http://localhost:8123/app/` to see the application. The API will be available at `http://localhost:8123`.
@@ -118,8 +113,7 @@ Open your browser and navigate to `http://localhost:8123/app/` to see the applic
 - [Tailwind CSS](https://tailwindcss.com/) - For styling.
 - [Shadcn UI](https://ui.shadcn.com/) - For components.
 - [LangGraph](https://github.com/langchain-ai/langgraph) - For building the backend research agent.
-- OpenAI-compatible chat model - LLM for query generation, reflection, and answer synthesis.
+- [Google Gemini](https://ai.google.dev/models/gemini) - LLM for query generation, reflection, and answer synthesis.
-- [Tavily](https://tavily.com/) - Web search for research retrieval.
 
 ## License
 

backend/.env.example

@@ -1 +1,3 @@
-# GEMINI_API_KEY=
+# OPENAI_API_KEY=
+# OPENAI_BASE_URL=
+# TAVILY_API_KEY=

backend/examples/cli_research.py

@@ -21,7 +21,7 @@ def main() -> None:
     )
     parser.add_argument(
         "--reasoning-model",
-        default="GLM-4.5-Air",
+        default="openai/gpt-5.4",
         help="Model for the final answer",
     )
     args = parser.parse_args()

backend/src/agent/configuration.py

@@ -9,26 +9,33 @@ class Configuration(BaseModel):
     """The configuration for the agent."""
 
     query_generator_model: str = Field(
-        default="GLM-4.5-Air",
+        default="openai/gpt-5.4",
         metadata={
             "description": "The name of the language model to use for the agent's query generation."
         },
     )
 
     reflection_model: str = Field(
-        default="GLM-4.5-Air",
+        default="openai/gpt-5.4",
         metadata={
             "description": "The name of the language model to use for the agent's reflection."
         },
     )
 
     answer_model: str = Field(
-        default="GLM-4.5-Air",
+        default="openai/gpt-5.4",
         metadata={
             "description": "The name of the language model to use for the agent's answer."
         },
     )
 
+    reasoning_model: str = Field(
+        default="openai/gpt-5.4",
+        metadata={
+            "description": "Fallback model used when the client does not provide a supported reasoning model."
+        },
+    )
+
     number_of_initial_queries: int = Field(
         default=3,
         metadata={"description": "The number of initial search queries to generate."},

backend/src/agent/graph.py

@@ -1,93 +1,134 @@
 import os
+from typing import Any
 
-from agent.tools_and_schemas import SearchQueryList, Reflection
+from agent.configuration import Configuration
-from dotenv import load_dotenv
+from agent.prompts import (
-from langchain_core.messages import AIMessage
+    answer_instructions,
-from langgraph.types import Send
+    get_current_date,
-from langgraph.graph import StateGraph
+    query_writer_instructions,
-from langgraph.graph import START, END
+    reflection_instructions,
-from langchain_core.runnables import RunnableConfig
+)
-from langchain_openai import ChatOpenAI
-from tavily import TavilyClient
-
 from agent.state import (
     OverallState,
     QueryGenerationState,
     ReflectionState,
     WebSearchState,
 )
-from agent.configuration import Configuration
+from agent.tools_and_schemas import Reflection, SearchQueryList
-from agent.prompts import (
-    get_current_date,
-    query_writer_instructions,
-    web_searcher_instructions,
-    reflection_instructions,
-    answer_instructions,
-)
 from agent.utils import (
-    deduplicate_and_format_sources,
+    format_sources_for_prompt,
     get_research_topic,
+    normalize_model_name,
+    normalize_tavily_sources,
+    shorten_search_query,
 )
+from dotenv import load_dotenv
+from langchain_core.messages import AIMessage
+from langchain_core.runnables import RunnableConfig
+from langchain_openai import ChatOpenAI
+from langgraph.graph import END, START, StateGraph
+from langgraph.types import Send
+from tavily import TavilyClient
 
 load_dotenv()
 
-if os.getenv("OPENAI_API_KEY") is None:
+TAVILY_TOPIC = "general"
-    raise ValueError("OPENAI_API_KEY is not set")
+TAVILY_SEARCH_DEPTH = "advanced"
-
+TAVILY_MAX_RESULTS = 5
-if os.getenv("TAVILY_API_KEY") is None:
+TAVILY_CHUNKS_PER_SOURCE = 3
-    raise ValueError("TAVILY_API_KEY is not set")
-
-tavily_client = TavilyClient(api_key=os.getenv("TAVILY_API_KEY"))
 
 
-def create_llm(model: str, temperature: float) -> ChatOpenAI:
+def require_env(name: str) -> str:
-    """Create an OpenAI-compatible chat client."""
+    """Return an environment variable or raise a clear error."""
+    value = os.getenv(name)
+    if not value:
+        raise ValueError(f"{name} is not set")
+    return value
+
+
+def get_chat_model(model_name: str, temperature: float = 0) -> ChatOpenAI:
+    """Create the OpenAI chat model lazily for easier testing."""
+    base_url = os.getenv("OPENAI_BASE_URL")
     return ChatOpenAI(
-        model=model,
+        model=normalize_model_name(model_name),
+        api_key=require_env("OPENAI_API_KEY"),
+        base_url=base_url or None,
         temperature=temperature,
-        max_retries=2,
-        api_key=os.getenv("OPENAI_API_KEY"),
-        base_url=os.getenv("OPENAI_BASE_URL", "https://open.bigmodel.cn/api/paas/v4"),
     )
 
 
-# Nodes
+def get_tavily_client() -> TavilyClient:
+    """Create the Tavily client lazily for easier testing."""
+    return TavilyClient(api_key=require_env("TAVILY_API_KEY"))
+
+
+def create_structured_response(
+    prompt: str,
+    model_name: str,
+    schema_model: type[Any],
+    temperature: float = 1,
+) -> Any:
+    """Call the LangChain OpenAI chat model and parse structured output."""
+    llm = get_chat_model(model_name, temperature=temperature)
+    return llm.with_structured_output(
+        schema_model,
+        method="function_calling",
+    ).invoke(prompt)
+
+
+def extract_text_content(content: Any) -> str:
+    """Normalize LangChain message content into a string."""
+    if isinstance(content, str):
+        return content.strip()
+
+    if isinstance(content, list):
+        text_chunks = []
+        for chunk in content:
+            if isinstance(chunk, dict) and chunk.get("type") == "text":
+                text_chunks.append(str(chunk.get("text", "")))
+        return "\n".join(text_chunks).strip()
+
+    return str(content).strip()
+
+
+def create_text_response(prompt: str, model_name: str, temperature: float = 0) -> str:
+    """Call the LangChain OpenAI chat model and return markdown output."""
+    response = get_chat_model(model_name, temperature=temperature).invoke(prompt)
+    text_output = extract_text_content(response.content)
+
+    if not text_output:
+        raise ValueError("OpenAI response did not include text output")
+
+    return text_output
+
+
 def generate_query(state: OverallState, config: RunnableConfig) -> QueryGenerationState:
-    """LangGraph node that generates search queries based on the User's question.
+    """Generate initial search queries from the user's request."""
-
-    Uses the configured chat model to create optimized search queries for web research.
-
-    Args:
-        state: Current graph state containing the User's question
-        config: Configuration for the runnable, including LLM provider settings
-
-    Returns:
-        Dictionary with state update, including search_query key containing the generated queries
-    """
     configurable = Configuration.from_runnable_config(config)
+    reasoning_model = normalize_model_name(
+        state.get("reasoning_model"),
+        configurable.reasoning_model or configurable.query_generator_model,
+    )
 
-    # Check for custom initial search query count.
     if state.get("initial_search_query_count") is None:
         state["initial_search_query_count"] = configurable.number_of_initial_queries
 
-    llm = create_llm(configurable.query_generator_model, temperature=1.0)
-    structured_llm = llm.with_structured_output(SearchQueryList)
-
-    current_date = get_current_date()
     formatted_prompt = query_writer_instructions.format(
-        current_date=current_date,
+        current_date=get_current_date(),
         research_topic=get_research_topic(state["messages"]),
         number_queries=state["initial_search_query_count"],
     )
-    result = structured_llm.invoke(formatted_prompt)
+    result = create_structured_response(
+        formatted_prompt,
+        reasoning_model,
+        SearchQueryList,
+        temperature=1,
+    )
     return {"search_query": result.query}
 
 
-def continue_to_web_research(state: QueryGenerationState):
+def continue_to_web_research(state: QueryGenerationState) -> list[Send]:
-    """LangGraph node that sends the search queries to the web research node.
+    """Fan out search queries into parallel web research nodes."""
-
-    This is used to spawn n number of web research nodes, one for each search query.
-    """
     return [
         Send("web_research", {"search_query": search_query, "id": int(idx)})
         for idx, search_query in enumerate(state["search_query"])
@@ -95,65 +136,63 @@ def continue_to_web_research(state: QueryGenerationState):
 
 
 def web_research(state: WebSearchState, config: RunnableConfig) -> OverallState:
-    """LangGraph node that performs web research using Tavily search."""
+    """Execute Tavily search and return raw evidence for one query."""
-    Configuration.from_runnable_config(config)
+    search_query = shorten_search_query(state["search_query"])
-    formatted_prompt = web_searcher_instructions.format(
+    tavily_response = get_tavily_client().search(
-        current_date=get_current_date(),
+        query=search_query,
-        research_topic=state["search_query"],
+        topic=TAVILY_TOPIC,
-    )
+        search_depth=TAVILY_SEARCH_DEPTH,
-    search_response = tavily_client.search(
+        chunks_per_source=TAVILY_CHUNKS_PER_SOURCE,
-        query=state["search_query"],
+        max_results=TAVILY_MAX_RESULTS,
-        topic="general",
-        search_depth="advanced",
-        max_results=5,
         include_answer=False,
         include_raw_content=False,
     )
-    sources_gathered, formatted_results = deduplicate_and_format_sources(
+    sources = normalize_tavily_sources(tavily_response.get("results", []))
-        search_response.get("results", []), int(state["id"])
-    )
 
-    web_research_result = (
+    if not sources:
-        f"{formatted_prompt}\n\n"
+        return {
-        f"Search query: {state['search_query']}\n\n"
+            "sources_gathered": [],
-        f"Findings:\n{formatted_results}"
+            "search_query": [search_query],
+            "web_research_result": [
+                f'No Tavily results were returned for "{search_query}".'
+            ],
+        }
+
+    evidence = "\n\n".join(
+        [
+            f"Search Query: {search_query}",
+            "Source Evidence:",
+            format_sources_for_prompt(sources),
+        ]
     )
 
     return {
-        "sources_gathered": sources_gathered,
+        "sources_gathered": sources,
-        "search_query": [state["search_query"]],
+        "search_query": [search_query],
-        "web_research_result": [web_research_result],
+        "web_research_result": [evidence],
     }
 
 
 def reflection(state: OverallState, config: RunnableConfig) -> ReflectionState:
-    """LangGraph node that identifies knowledge gaps and generates potential follow-up queries.
+    """Decide whether additional research is needed."""
-
-    Analyzes the current summary to identify areas for further research and generates
-    potential follow-up queries. Uses structured output to extract
-    the follow-up query in JSON format.
-
-    Args:
-        state: Current graph state containing the running summary and research topic
-        config: Configuration for the runnable, including LLM provider settings
-
-    Returns:
-        Dictionary with state update, including search_query key containing the generated follow-up query
-    """
     configurable = Configuration.from_runnable_config(config)
-    # Increment the research loop count and get the reasoning model
     state["research_loop_count"] = state.get("research_loop_count", 0) + 1
-    reasoning_model = state.get("reasoning_model", configurable.reflection_model)
+    reasoning_model = normalize_model_name(
+        state.get("reasoning_model"),
+        configurable.reasoning_model,
+    )
 
-    # Format the prompt
-    current_date = get_current_date()
     formatted_prompt = reflection_instructions.format(
-        current_date=current_date,
+        current_date=get_current_date(),
         research_topic=get_research_topic(state["messages"]),
         summaries="\n\n---\n\n".join(state["web_research_result"]),
     )
-    llm = create_llm(reasoning_model, temperature=1.0)
+    result = create_structured_response(
-    result = llm.with_structured_output(Reflection).invoke(formatted_prompt)
+        formatted_prompt,
+        reasoning_model,
+        Reflection,
+        temperature=1,
+    )
 
     return {
         "is_sufficient": result.is_sufficient,
@@ -167,19 +206,8 @@ def reflection(state: OverallState, config: RunnableConfig) -> ReflectionState:
 def evaluate_research(
     state: ReflectionState,
     config: RunnableConfig,
-) -> OverallState:
+) -> OverallState | list[Send]:
-    """LangGraph routing function that determines the next step in the research flow.
+    """Route either back into research or into final answer synthesis."""
-
-    Controls the research loop by deciding whether to continue gathering information
-    or to finalize the summary based on the configured maximum number of research loops.
-
-    Args:
-        state: Current graph state containing the research loop count
-        config: Configuration for the runnable, including max_research_loops setting
-
-    Returns:
-        String literal indicating the next node to visit ("web_research" or "finalize_summary")
-    """
     configurable = Configuration.from_runnable_config(config)
     max_research_loops = (
         state.get("max_research_loops")
@@ -188,7 +216,7 @@ def evaluate_research(
     )
     if state["is_sufficient"] or state["research_loop_count"] >= max_research_loops:
         return "finalize_answer"
-    else:
+
     return [
         Send(
             "web_research",
@@ -201,71 +229,40 @@ def evaluate_research(
     ]
 
 
-def finalize_answer(state: OverallState, config: RunnableConfig):
+def finalize_answer(state: OverallState, config: RunnableConfig) -> OverallState:
-    """LangGraph node that finalizes the research summary.
+    """Generate the final cited answer from the accumulated research summaries."""
-
-    Prepares the final output by deduplicating and formatting sources, then
-    combining them with the running summary to create a well-structured
-    research report with proper citations.
-
-    Args:
-        state: Current graph state containing the running summary and sources gathered
-
-    Returns:
-        Dictionary with state update, including running_summary key containing the formatted final summary with sources
-    """
     configurable = Configuration.from_runnable_config(config)
-    reasoning_model = state.get("reasoning_model") or configurable.answer_model
+    reasoning_model = normalize_model_name(
+        state.get("reasoning_model"),
+        configurable.reasoning_model or configurable.answer_model,
+    )
 
-    # Format the prompt
-    current_date = get_current_date()
     formatted_prompt = answer_instructions.format(
-        current_date=current_date,
+        current_date=get_current_date(),
         research_topic=get_research_topic(state["messages"]),
         summaries="\n---\n\n".join(state["web_research_result"]),
     )
-
+    answer = create_text_response(formatted_prompt, reasoning_model, temperature=0)
-    llm = create_llm(reasoning_model, temperature=0)
-    result = llm.invoke(formatted_prompt)
-
-    # Replace the short urls with the original urls and add all used urls to the sources_gathered
-    unique_sources = []
-    for source in state["sources_gathered"]:
-        if source["short_url"] in result.content:
-            result.content = result.content.replace(
-                source["short_url"], source["value"]
-            )
-            unique_sources.append(source)
 
     return {
-        "messages": [AIMessage(content=result.content)],
+        "messages": [AIMessage(content=answer)],
-        "sources_gathered": unique_sources,
+        "sources_gathered": state.get("sources_gathered", []),
     }
 
 
-# Create our Agent Graph
 builder = StateGraph(OverallState, config_schema=Configuration)
-
-# Define the nodes we will cycle between
 builder.add_node("generate_query", generate_query)
 builder.add_node("web_research", web_research)
 builder.add_node("reflection", reflection)
 builder.add_node("finalize_answer", finalize_answer)
-
-# Set the entrypoint as `generate_query`
-# This means that this node is the first one called
 builder.add_edge(START, "generate_query")
-# Add conditional edge to continue with search queries in a parallel branch
 builder.add_conditional_edges(
     "generate_query", continue_to_web_research, ["web_research"]
 )
-# Reflect on the web research
 builder.add_edge("web_research", "reflection")
-# Evaluate the research
 builder.add_conditional_edges(
     "reflection", evaluate_research, ["web_research", "finalize_answer"]
 )
-# Finalize the answer
 builder.add_edge("finalize_answer", END)
 
 graph = builder.compile(name="pro-search-agent")

backend/src/agent/prompts.py

@@ -15,6 +15,7 @@ Instructions:
 - Queries should be diverse, if the topic is broad, generate more than 1 query.
 - Don't generate multiple similar queries, 1 is enough.
 - Query should ensure that the most current information is gathered. The current date is {current_date}.
+- Keep every individual query under 400 characters. Prefer concise search terms over full-sentence restatements.
 
 Format: 
 - Format your response as a JSON object with ALL two of these exact keys:
@@ -34,26 +35,31 @@ Topic: What revenue grew more last year apple stock or the number of people buyi
 Context: {research_topic}"""
 
 
-web_searcher_instructions = """Conduct targeted web research to gather the most recent, credible information on "{research_topic}" and synthesize it into a verifiable text artifact.
+web_searcher_instructions = """Review the provided Tavily search results for "{research_topic}" and synthesize them into a verifiable research note.
 
 Instructions:
 - Query should ensure that the most current information is gathered. The current date is {current_date}.
-- Review the returned web search results and extract the strongest, most relevant findings.
+- Use only the provided Tavily search results. Do not invent or infer facts that are not supported by those results.
-- Consolidate key findings while meticulously tracking the source(s) for each specific piece of information.
+- Consolidate the key findings into a concise research note for this query.
-- The output should be a well-written summary or report based on your search findings. 
+- Every factual paragraph or bullet must include at least one markdown citation using the exact title and URL from the provided sources, for example: [Reuters](https://www.reuters.com/example).
-- Only include the information found in the search results, don't make up any information.
+- Preserve full URLs in citations. Do not use placeholders, IDs, or shortened URLs.
+- If the results are insufficient, explicitly say what is missing and cite the closest available sources.
 
 Research Topic:
 {research_topic}
+
+Search Results:
+{search_results}
 """
 
-reflection_instructions = """You are an expert research assistant analyzing summaries about "{research_topic}".
+reflection_instructions = """You are an expert research assistant analyzing collected research evidence about "{research_topic}".
 
 Instructions:
 - Identify knowledge gaps or areas that need deeper exploration and generate a follow-up query. (1 or multiple).
-- If provided summaries are sufficient to answer the user's question, don't generate a follow-up query.
+- If the provided evidence is sufficient to answer the user's question, don't generate a follow-up query.
 - If there is a knowledge gap, generate a follow-up query that would help expand your understanding.
 - Focus on technical details, implementation specifics, or emerging trends that weren't fully covered.
+- Keep every follow-up query under 400 characters. Prefer compact search phrases over long natural-language questions.
 
 Requirements:
 - Ensure the follow-up query is self-contained and includes necessary context for web search.
@@ -73,24 +79,26 @@ Example:
 }}
 ```
 
-Reflect carefully on the Summaries to identify knowledge gaps and produce a follow-up query. Then, produce your output following this JSON format:
+Reflect carefully on the Research Evidence to identify knowledge gaps and produce a follow-up query. Then, produce your output following this JSON format:
 
-Summaries:
+Research Evidence:
 {summaries}
 """
 
-answer_instructions = """Generate a high-quality answer to the user's question based on the provided summaries.
+answer_instructions = """Generate a high-quality answer to the user's question based on the provided research evidence.
 
 Instructions:
 - The current date is {current_date}.
 - You are the final step of a multi-step research process, don't mention that you are the final step. 
 - You have access to all the information gathered from the previous steps.
 - You have access to the user's question.
-- Generate a high-quality answer to the user's question based on the provided summaries and the user's question.
+- Generate a high-quality answer to the user's question based on the provided research evidence and the user's question.
-- Include the sources you used from the Summaries in the answer correctly, using markdown links. THIS IS A MUST.
+- Use only the evidence present in the Research Evidence.
+- Preserve or reuse source citations from the Research Evidence as full markdown links. Every factual paragraph should include at least one citation.
+- Do not invent sources, placeholders, or shortened URLs.
 
 User Context:
 - {research_topic}
 
-Summaries:
+Research Evidence:
 {summaries}"""

backend/src/agent/state.py

@@ -29,13 +29,8 @@ class ReflectionState(TypedDict):
     number_of_ran_queries: int
 
 
-class Query(TypedDict):
-    query: str
-    rationale: str
-
-
 class QueryGenerationState(TypedDict):
-    search_query: list[Query]
+    search_query: list[str]
 
 
 class WebSearchState(TypedDict):

backend/src/agent/utils.py

@@ -1,9 +1,16 @@
-from typing import Any
+from __future__ import annotations
+
+import re
+from typing import Any, List
 
 from langchain_core.messages import AIMessage, AnyMessage, HumanMessage
 
+DEFAULT_REASONING_MODEL = "openai/gpt-5.4"
+SUPPORTED_MODEL_PREFIXES = ("gpt-", "deepseek-", "glm-")
+MAX_TAVILY_QUERY_LENGTH = 380
 
-def get_research_topic(messages: list[AnyMessage]) -> str:
+
+def get_research_topic(messages: List[AnyMessage]) -> str:
     """Get the research topic from the messages."""
     if len(messages) == 1:
         return str(messages[-1].content)
@@ -17,43 +24,83 @@ def get_research_topic(messages: list[AnyMessage]) -> str:
     return research_topic
 
 
-def deduplicate_and_format_sources(
+def normalize_model_name(
-    results: list[dict[str, Any]], search_id: int
+    model_name: str | None,
-) -> tuple[list[dict[str, str]], str]:
+    fallback: str = DEFAULT_REASONING_MODEL,
-    """Convert Tavily results into source metadata and a markdown research summary."""
+) -> str:
-    sources: list[dict[str, str]] = []
+    """Normalize stale or unsupported model names to a supported default."""
-    formatted_sections: list[str] = []
+    candidate = (model_name or "").strip()
+    if candidate.startswith(SUPPORTED_MODEL_PREFIXES):
+        return candidate
+
+    provider, _, model = candidate.partition("/")
+    if provider and model.startswith(SUPPORTED_MODEL_PREFIXES):
+        return candidate
+
+    return fallback
+
+
+def _normalize_source_title(source: dict[str, Any], index: int) -> str:
+    title = (source.get("title") or "").strip()
+    return title or f"Source {index}"
+
+
+def normalize_tavily_sources(results: list[dict[str, Any]]) -> list[dict[str, str]]:
+    """Normalize Tavily results into a stable source shape."""
+    normalized_sources: list[dict[str, str]] = []
     seen_urls: set[str] = set()
 
-    for idx, result in enumerate(results):
+    for index, result in enumerate(results, start=1):
-        url = result.get("url")
+        url = (result.get("url") or "").strip()
         if not url or url in seen_urls:
             continue
 
-        seen_urls.add(url)
+        normalized_sources.append(
-        title = result.get("title") or f"Source {idx + 1}"
-        short_url = f"https://search.local/{search_id}-{idx}"
-        content = (result.get("content") or "").strip()
-
-        sources.append(
             {
-                "label": title,
+                "label": f"S{len(normalized_sources) + 1}",
-                "short_url": short_url,
+                "title": _normalize_source_title(result, index),
                 "value": url,
+                "content": (result.get("content") or "").strip(),
             }
         )
+        seen_urls.add(url)
 
-        formatted_sections.append(
+    return normalized_sources
+
+
+def format_sources_for_prompt(sources: list[dict[str, str]]) -> str:
+    """Format normalized search results for prompt injection."""
+    if not sources:
+        return "No search results were returned."
+
+    formatted_sources = []
+    for source in sources:
+        snippet = source.get("content") or "No snippet available."
+        formatted_sources.append(
             "\n".join(
                 [
-                    f"- {title} [{title}]({short_url})",
+                    f"{source['label']}: {source['title']}",
-                    f"  URL: {url}",
+                    f"URL: {source['value']}",
-                    f"  Snippet: {content or 'No snippet returned.'}",
+                    f"Snippet: {snippet}",
                 ]
             )
         )
 
-    if not formatted_sections:
+    return "\n\n".join(formatted_sources)
-        formatted_sections.append("- No Tavily results were returned for this query.")
 
-    return sources, "\n".join(formatted_sections)
+
+def shorten_search_query(
+    query: str,
+    max_length: int = MAX_TAVILY_QUERY_LENGTH,
+) -> str:
+    """Normalize and trim search queries to fit Tavily's length limit."""
+    normalized_query = re.sub(r"\s+", " ", query).strip()
+    if len(normalized_query) <= max_length:
+        return normalized_query
+
+    truncated_query = normalized_query[:max_length]
+    last_space = truncated_query.rfind(" ")
+    if last_space > max_length // 2:
+        truncated_query = truncated_query[:last_space]
+
+    return truncated_query.rstrip(" ,;:-")

docker-compose.yml

@@ -28,7 +28,7 @@ services:
       retries: 5
       interval: 5s
   langgraph-api:
-    image: openai-fullstack-langgraph
+    image: gemini-fullstack-langgraph
     container_name: langgraph-api
     ports:
       - "8123:8000"
@@ -39,7 +39,7 @@ services:
         condition: service_healthy
     environment:
       OPENAI_API_KEY: ${OPENAI_API_KEY}
-      OPENAI_BASE_URL: ${OPENAI_BASE_URL:-https://open.bigmodel.cn/api/paas/v4}
+      OPENAI_BASE_URL: ${OPENAI_BASE_URL}
       TAVILY_API_KEY: ${TAVILY_API_KEY}
       LANGSMITH_API_KEY: ${LANGSMITH_API_KEY}
       REDIS_URI: redis://langgraph-redis:6379

frontend/package-lock.json

@@ -2259,7 +2259,6 @@
       "integrity": "sha512-wIX2aSZL5FE+MR0JlvF87BNVrtFWf6AE6rxSE9X7OwnVvoyCQjpzSRJ+M87se/4QCkCiebQAqrJ0y6fwIyi7nw==",
       "devOptional": true,
       "license": "MIT",
-      "peer": true,
       "dependencies": {
         "undici-types": "~6.21.0"
       }
@@ -2269,7 +2268,6 @@
       "resolved": "https://registry.npmjs.org/@types/react/-/react-19.1.2.tgz",
       "integrity": "sha512-oxLPMytKchWGbnQM9O7D67uPa9paTNxO7jVoNMXgkkErULBPhPARCfkKL9ytcIJJRGjbsVwW4ugJzyFFvm/Tiw==",
       "license": "MIT",
-      "peer": true,
       "dependencies": {
         "csstype": "^3.0.2"
       }
@@ -2280,7 +2278,6 @@
       "integrity": "sha512-rJXC08OG0h3W6wDMFxQrZF00Kq6qQvw0djHRdzl3U5DnIERz0MRce3WVc7IS6JYBwtaP/DwYtRRjVlvivNveKg==",
       "devOptional": true,
       "license": "MIT",
-      "peer": true,
       "peerDependencies": {
         "@types/react": "^19.0.0"
       }
@@ -2339,7 +2336,6 @@
       "integrity": "sha512-oU/OtYVydhXnumd0BobL9rkJg7wFJ9bFFPmSmB/bf/XWN85hlViji59ko6bSKBXyseT9V8l+CN1nwmlbiN0G7Q==",
       "dev": true,
       "license": "MIT",
-      "peer": true,
       "dependencies": {
         "@typescript-eslint/scope-manager": "8.31.1",
         "@typescript-eslint/types": "8.31.1",
@@ -2535,7 +2531,6 @@
       "integrity": "sha512-OvQ/2pUDKmgfCg++xsTX1wGxfTaszcHVcTctW4UJB4hibJx2HXxxO5UmVgyjMa+ZDsiaf5wWLXYpRWMmBI0QHg==",
       "dev": true,
       "license": "MIT",
-      "peer": true,
       "bin": {
         "acorn": "bin/acorn"
       },
@@ -3003,7 +2998,6 @@
       "integrity": "sha512-E6Mtz9oGQWDCpV12319d59n4tx9zOTXSTmc8BLVxBx+G/0RdM5MvEEJLU9c0+aleoePYYgVTOsRblx433qmhWQ==",
       "dev": true,
       "license": "MIT",
-      "peer": true,
       "dependencies": {
         "@eslint-community/eslint-utils": "^4.2.0",
         "@eslint-community/regexpp": "^4.12.1",
@@ -4885,7 +4879,6 @@
       "resolved": "https://registry.npmjs.org/react/-/react-19.1.0.tgz",
       "integrity": "sha512-FS+XFBNvn3GTAWq26joslQgWNoFu08F4kl0J4CgdNKADkdSGXQyTCnKteIAJy96Br6YbpEU1LSzV5dYtjMkMDg==",
       "license": "MIT",
-      "peer": true,
       "engines": {
         "node": ">=0.10.0"
       }
@@ -4895,7 +4888,6 @@
       "resolved": "https://registry.npmjs.org/react-dom/-/react-dom-19.1.0.tgz",
       "integrity": "sha512-Xs1hdnE+DyKgeHJeJznQmYMIBG3TKIHJJT95Q58nHLSrElKlGQqDTR2HQ9fx5CN/Gk6Vh/kupBTDLU11/nDk/g==",
       "license": "MIT",
-      "peer": true,
       "dependencies": {
         "scheduler": "^0.26.0"
       },
@@ -5353,7 +5345,6 @@
       "resolved": "https://registry.npmjs.org/picomatch/-/picomatch-4.0.2.tgz",
       "integrity": "sha512-M7BAV6Rlcy5u+m6oPhAPFgJTzAioX/6B0DxyvDlo9l8+T3nLKbrczg2WLUyzd45L8RqfUMyGPzekbMvX2Ldkwg==",
       "license": "MIT",
-      "peer": true,
       "engines": {
         "node": ">=12"
       },
@@ -5448,7 +5439,6 @@
       "integrity": "sha512-84MVSjMEHP+FQRPy3pX9sTVV/INIex71s9TL2Gm5FG/WG1SqXeKyZ0k7/blY/4FdOzI12CBy1vGc4og/eus0fw==",
       "dev": true,
       "license": "Apache-2.0",
-      "peer": true,
       "bin": {
         "tsc": "bin/tsc",
         "tsserver": "bin/tsserver"
@@ -5673,7 +5663,6 @@
       "resolved": "https://registry.npmjs.org/vite/-/vite-6.3.4.tgz",
       "integrity": "sha512-BiReIiMS2fyFqbqNT/Qqt4CVITDU9M9vE+DKcVAsB+ZV0wvTKd+3hMbkpxz1b+NmEDMegpVbisKiAZOnvO92Sw==",
       "license": "MIT",
-      "peer": true,
       "dependencies": {
         "esbuild": "^0.25.0",
         "fdir": "^6.4.4",
@@ -5762,7 +5751,6 @@
       "resolved": "https://registry.npmjs.org/picomatch/-/picomatch-4.0.2.tgz",
       "integrity": "sha512-M7BAV6Rlcy5u+m6oPhAPFgJTzAioX/6B0DxyvDlo9l8+T3nLKbrczg2WLUyzd45L8RqfUMyGPzekbMvX2Ldkwg==",
       "license": "MIT",
-      "peer": true,
       "engines": {
         "node": ">=12"
       },
@@ -5814,7 +5802,6 @@
       "resolved": "https://registry.npmjs.org/zod/-/zod-3.24.4.tgz",
       "integrity": "sha512-OdqJE9UDRPwWsrHjLN2F8bPxvwJBK22EHLWtanu0LSYr5YqzsaaW3RMgmjwr8Rypg5k+meEJdSPXJZXE/yqOMg==",
       "license": "MIT",
-      "peer": true,
       "funding": {
         "url": "https://github.com/sponsors/colinhacks"
       }

frontend/src/App.tsx

@@ -6,10 +6,27 @@ import { WelcomeScreen } from "@/components/WelcomeScreen";
 import { ChatMessagesView } from "@/components/ChatMessagesView";
 import { Button } from "@/components/ui/button";
 
+const getSourceDisplayName = (source: any): string | null => {
+  if (source?.title) {
+    return source.title;
+  }
+
+  if (source?.value) {
+    try {
+      return new URL(source.value).hostname.replace(/^www\./, "");
+    } catch {
+      return null;
+    }
+  }
+
+  return source?.label || null;
+};
+
 export default function App() {
   const [processedEventsTimeline, setProcessedEventsTimeline] = useState<
     ProcessedEvent[]
   >([]);
+  const [selectedModel, setSelectedModel] = useState("openai/gpt-5.4");
   const [historicalActivities, setHistoricalActivities] = useState<
     Record<string, ProcessedEvent[]>
   >({});
@@ -38,7 +55,7 @@ export default function App() {
         const sources = event.web_research.sources_gathered || [];
         const numSources = sources.length;
         const uniqueLabels = [
-          ...new Set(sources.map((s: any) => s.label).filter(Boolean)),
+          ...new Set(sources.map(getSourceDisplayName).filter(Boolean)),
         ];
         const exampleLabels = uniqueLabels.slice(0, 3).join(", ");
         processedEvent = {
@@ -157,6 +174,8 @@ export default function App() {
               handleSubmit={handleSubmit}
               isLoading={thread.isLoading}
               onCancel={handleCancel}
+              model={selectedModel}
+              onModelChange={setSelectedModel}
             />
           ) : error ? (
             <div className="flex flex-col items-center justify-center h-full">
@@ -181,6 +200,8 @@ export default function App() {
               onCancel={handleCancel}
               liveActivityEvents={processedEventsTimeline}
               historicalActivities={historicalActivities}
+              model={selectedModel}
+              onModelChange={setSelectedModel}
             />
           )}
       </main>

frontend/src/components/ChatMessagesView.tsx

@@ -230,6 +230,8 @@ interface ChatMessagesViewProps {
   onCancel: () => void;
   liveActivityEvents: ProcessedEvent[];
   historicalActivities: Record<string, ProcessedEvent[]>;
+  model: string;
+  onModelChange: (model: string) => void;
 }
 
 export function ChatMessagesView({
@@ -240,6 +242,8 @@ export function ChatMessagesView({
   onCancel,
   liveActivityEvents,
   historicalActivities,
+  model,
+  onModelChange,
 }: ChatMessagesViewProps) {
   const [copiedMessageId, setCopiedMessageId] = useState<string | null>(null);
 
@@ -316,6 +320,8 @@ export function ChatMessagesView({
         isLoading={isLoading}
         onCancel={onCancel}
         hasHistory={messages.length > 0}
+        model={model}
+        onModelChange={onModelChange}
       />
     </div>
   );

frontend/src/components/InputForm.tsx

@@ -16,6 +16,8 @@ interface InputFormProps {
   onCancel: () => void;
   isLoading: boolean;
   hasHistory: boolean;
+  model: string;
+  onModelChange: (model: string) => void;
 }
 
 export const InputForm: React.FC<InputFormProps> = ({
@@ -23,10 +25,11 @@ export const InputForm: React.FC<InputFormProps> = ({
   onCancel,
   isLoading,
   hasHistory,
+  model,
+  onModelChange,
 }) => {
   const [internalInputValue, setInternalInputValue] = useState("");
   const [effort, setEffort] = useState("medium");
-  const [model, setModel] = useState("GLM-4.5-Air");
 
   const handleInternalSubmit = (e?: React.FormEvent) => {
     if (e) e.preventDefault();
@@ -130,17 +133,41 @@ export const InputForm: React.FC<InputFormProps> = ({
               <Cpu className="h-4 w-4 mr-2" />
               Model
             </div>
-            <Select value={model} onValueChange={setModel}>
+            <Select value={model} onValueChange={onModelChange}>
               <SelectTrigger className="w-[150px] bg-transparent border-none cursor-pointer">
                 <SelectValue placeholder="Model" />
               </SelectTrigger>
               <SelectContent className="bg-neutral-700 border-neutral-600 text-neutral-300 cursor-pointer">
                 <SelectItem
-                  value="GLM-4.5-Air"
+                  value="openai/gpt-5.4"
                   className="hover:bg-neutral-600 focus:bg-neutral-600 cursor-pointer"
                 >
                   <div className="flex items-center">
-                    <Zap className="h-4 w-4 mr-2 text-cyan-400" /> GLM-4.5-Air
+                    <Zap className="h-4 w-4 mr-2 text-emerald-400" /> openai/gpt-5.4
+                  </div>
+                </SelectItem>
+                <SelectItem
+                  value="deepseek-v3"
+                  className="hover:bg-neutral-600 focus:bg-neutral-600 cursor-pointer"
+                >
+                  <div className="flex items-center">
+                    <Cpu className="h-4 w-4 mr-2 text-cyan-400" /> deepseek-v3
+                  </div>
+                </SelectItem>
+                <SelectItem
+                  value="glm-4.5-air"
+                  className="hover:bg-neutral-600 focus:bg-neutral-600 cursor-pointer"
+                >
+                  <div className="flex items-center">
+                    <Cpu className="h-4 w-4 mr-2 text-sky-400" /> glm-4.5-air
+                  </div>
+                </SelectItem>
+                <SelectItem
+                  value="glm-4.7"
+                  className="hover:bg-neutral-600 focus:bg-neutral-600 cursor-pointer"
+                >
+                  <div className="flex items-center">
+                    <Cpu className="h-4 w-4 mr-2 text-violet-400" /> glm-4.7
                   </div>
                 </SelectItem>
               </SelectContent>

frontend/src/components/WelcomeScreen.tsx

@@ -8,12 +8,16 @@ interface WelcomeScreenProps {
   ) => void;
   onCancel: () => void;
   isLoading: boolean;
+  model: string;
+  onModelChange: (model: string) => void;
 }
 
 export const WelcomeScreen: React.FC<WelcomeScreenProps> = ({
   handleSubmit,
   onCancel,
   isLoading,
+  model,
+  onModelChange,
 }) => (
   <div className="h-full flex flex-col items-center justify-center text-center px-4 flex-1 w-full max-w-3xl mx-auto gap-4">
     <div>
@@ -30,10 +34,12 @@ export const WelcomeScreen: React.FC<WelcomeScreenProps> = ({
         isLoading={isLoading}
         onCancel={onCancel}
         hasHistory={false}
+        model={model}
+        onModelChange={onModelChange}
       />
     </div>
     <p className="text-xs text-neutral-500">
-      Powered by GLM-4.5-Air, Tavily, and LangChain LangGraph.
+      Powered by Google Gemini and LangChain LangGraph.
     </p>
   </div>
 );