SummarizationNode: Please provide the conversation

[Kamran Huseynov (Kamran-ov)]

Hi,

I am having some issues on Summarization, and hope you can help me to fix this issue. For some reason, my agent cannot take the messages to summarize them before my second invoke.
Last output: "Please provide the conversation you would like me to summarize"

I am new to building AI Agents, so I would appreciate your help on this matter.

from typing import List, Any
from langmem.short_term import SummarizationNode
from langchain_core.messages.utils import count_tokens_approximately
from langgraph.prebuilt.chat_agent_executor import AgentState
from langgraph.prebuilt import create_react_agent
from langgraph.checkpoint.memory import MemorySaver
from langchain_google_genai import ChatGoogleGenerativeAI
from langchain_core.prompts import ChatPromptTemplate, MessagesPlaceholder
from langchain_community.tools.tavily_search import TavilySearchResults

class State(AgentState):
    # NOTE: we're adding this key to keep track of previous summary information
    # to make sure we're not summarizing on every LLM call
    context: dict[str, Any]
    input: str 


# 1. Create Model
llm = ChatGoogleGenerativeAI(
    model="gemini-2.0-flash",
    google_api_key=ai_api_key,
    temperature=0.7,
)

# 2. Create PROPER prompt template
prompt = ChatPromptTemplate.from_messages([
    ("system", "\n".join(system_prompt)),
    MessagesPlaceholder(variable_name="messages"),
    ("human", "{input}"),
    MessagesPlaceholder(variable_name="llm_input_messages"),
])

# 3. Configure Tavily Search tool
tavily_tool = [TavilySearchResults(
    name="food_research",
    description=(
        "Specialized search for food science, nutrition studies, "
        "and ingredient research. Use for factual verification and "
        "scientific context about food products."
    ),
    max_results=3,
    include_raw_content=True
)]

# 4 Create Summarization logic
summarization_node = SummarizationNode(
    token_counter=count_tokens_approximately,
    model=llm,
    max_tokens=384,
    max_summary_tokens=128,
    output_messages_key="llm_input_messages",
)

# 5. Create Memory Saver Checkpointer
checkpointer = MemorySaver()

# 6. Create agent with exact parameters
agent = create_react_agent(
    model=llm,
    tools=tavily_tool,
    pre_model_hook=summarization_node,
    state_schema=State,
    prompt=prompt,
    checkpointer=checkpointer,
    version="v2",  
    name="food_product_analyzer",
)

Invoke 1:

thread_id = "thread-12345"
current_state = {
    "input": "What is the nutritional value of Spirulina? ",
}
config = {"configurable": {"thread_id": thread_id}}
result = agent.invoke(current_state, config)

Output 1:
{'messages': [AIMessage(content="html\n <p>Here's a breakdown of the nutritional value of Spirulina:</p>\n <hr>\n <strong>🔵 BASIC: Spirulina is like a super-powered veggie from the water! </strong> It's packed with vitamins and protein, like a tiny pill that gives you a boost.\n <hr>\n <strong> 🟢 DETAILED: Nutritional Profile of Spirulina </strong>\n <ul>\n <li><strong>Protein:</strong> Spirulina is a complete protein source, containing all essential amino acids.</li>\n <li><strong>Vitamins:</strong> Rich in B vitamins (B1, B2, B3), vitamin E, and vitamin K.</li>\n <li><strong>Minerals:</strong> Contains iron, copper, magnesium, potassium, and manganese.</li>\n <li><strong>Antioxidants:</strong> High in phycocyanin, a powerful antioxidant with anti-inflammatory properties.</li>\n <li><strong>Other Nutrients:</strong> Includes essential fatty acids like gamma-linolenic acid (GLA) and dietary fiber.</li>\n </ul>\n <hr>\n <strong> 🔴 EXPERT: In-depth Nutritional and Biochemical Aspects of Spirulina </strong>\n <p>Spirulina's high protein content (around 60-70% by dry weight) features all essential amino acids, making it comparable to eggs. Its vitamin B complex supports energy metabolism, while vitamin E acts as a lipid-soluble antioxidant, protecting cell membranes from oxidative damage[^1]. Key minerals like iron contribute to oxygen transport and enzymatic functions. Phycocyanin, a unique biliprotein, gives Spirulina its blue-green color and exhibits potent antioxidant and anti-inflammatory effects by inhibiting reactive oxygen species and modulating inflammatory pathways such as the NF-κB pathway[^2]. Furthermore, gamma-linolenic acid (GLA) supports healthy inflammatory response.</p>\n <hr>\n 📚 Scientific References\n <p>[^1]: Amaro, H. M., Guedes, A. C., & Malcata, F. X. (2011). Advances in the production of high-value-added compounds using microalgae: a review. Journal of Chemical Technology & Biotechnology, 86(10), 1271-1289.</p>\n <p>[^2]: Eriksen, N. T. (2008). Production of phycocyanin: a review. Journal of Applied Phycology, 20(3), 201-211.</p>\n ", additional_kwargs={}, response_metadata={'prompt_feedback': {'block_reason': 0, 'safety_ratings': []}, 'finish_reason': 'STOP', 'model_name': 'gemini-2.0-flash', 'safety_ratings': []}, name='food_product_analyzer', id='run--f586bc4a-6e68-4a46-ba67-a78650ef92ca-0', usage_metadata={'input_tokens': 267, 'output_tokens': 538, 'total_tokens': 805, 'input_token_details': {'cache_read': 0}})],
'input': 'What is the nutritional value of Spirulina?. '}

Invoke 2:

current_state = {
    "input": "What are the health benefits of it? ",
}
result = agent.invoke(current_state, config)

Output 2:
{'messages': [

AIMessage(Here it is just first invoke response),

AIMessage(content="Please provide me with the name of the food or ingredient you're asking about so I can look up its health benefits.", additional_kwargs={}, response_metadata={'prompt_feedback': {'block_reason': 0, 'safety_ratings': []}, 'finish_reason': 'STOP', 'model_name': 'gemini-2.0-flash', 'safety_ratings': []}, name='food_product_analyzer', id='run--397ec349-45e6-4bc4-b143-1c2f6cd258e5-0', usage_metadata={'input_tokens': 332, 'output_tokens': 26, 'total_tokens': 358, 'input_token_details': {'cache_read': 0}})],

'context': {'running_summary': RunningSummary(summary='Please provide the conversation you would like me to summarize. I need the text of the conversation to be able to create a summary.', summarized_message_ids={'run--f586bc4a-6e68-4a46-ba67-a78650ef92ca-0'}, last_summarized_message_id='run--f586bc4a-6e68-4a46-ba67-a78650ef92ca-0')},

'input': 'What are the health benefits of it? '}

[jmaillefaud]

This is a problematic interaction with create_react_executor. In the call_model node, create_react_executor will extract the llm_input_messages key and override messages with it. When the summarization node is reached again, there no differences between the llm_input_messages key (containing the summarized messages) and the messages key (containing all the messages). As SummarizationNode will remove equal messages, there are no messages to summarize.