All projects & work

Context-Aware Agent

A local ReAct assistant that checks context and chooses tools.

Personal project

Context tools
4
ReAct iteration limit
5
Interfaces
Flask + CLI
Default Ollama model
llama3
Context-Aware Agent workflow diagram
Workflow diagram based on the repository’s implementation.

LangChain · Ollama · Flask

View repository

The idea

I built this assistant around a small question: when I paste background into a chat, how does the model decide which parts to use? I gave the Flask interface and the terminal loop access to the same local model so I could try the same question through either one.

How a question reaches the local model

I can ask through Flask or the CLI. Both feed the same runner, and the browser has an additional recovery path.Flask chatCLIReAct runnerContext toolsSearch toolModel response
I can ask through Flask or the CLI. Both feed the same runner, and the browser has an additional recovery path.

What happens when the answer falls short

I gave the runner four tools. A question with pasted notes can use the context checks. A question that needs outside information can use search. The model decides which tools to call.

  • Wikipedia is the default search source.
  • The executor stops after five iterations.
  • Flask can try the smart and manual workflows if the first answer fails its checks.

A closer look at the code

Choose and probe the local model

This excerpt selects the configured Ollama model and checks the connection during initialization. The surrounding function reports and re-raises initialization errors.

model_name = os.getenv("OLLAMA_MODEL", "llama3")
llm = OllamaLLM(model=model_name)

# Test the connection
test_response = llm.invoke("Hello")  # Updated to use invoke
return llm

Scope & practical notes

  • A relevance score or a fluent answer still needs a human check.