New ask Command Shipping with v1.5

[logan-markewich]

v1.5 is shipping with a new ask command in the CLI. This is essentially a RAG agent that will search your files and cite it's sources!

# Perform agentic RAG search
$ ask "What papers discuss graph neural networks?" dir/*.txt 

# Ask over content without agentic RAG
$ cat *.txt | ask "Summarize this"

# Ask over PDFs/unstructured files
$ parse dir/*.pdf | xargs ask "Some query"

This is the same syntax as the search tool, but instead, uses agentic RAG to search your files or respond directly

How the Agent Works

The agent has 3 main tools

• search -- this is just the search CLI command packaged as an agent tool. It works the same way with the same options, and also integrates with workspaces
• grep -- a standard grep tool for exact pattern matching. This building on top of the grep crate built by the ripgrep team
• read -- a tool to open and read a file, with options for specifying line numbers

All three of these tools have the same return structure, so the agent can deal with a familiar interface across all 3 tools.

Here's a sample of the returned format:

<chunk file={filename} start={start} end={end}>
{chunk_text}
</chunk>

stdin Mode

The one caveat here is when piping in content directly with stdin, all agent tools are removed, and the LLM responds directly using the provided text as context.

Changing the LLM

By default, ask uses gpt-4.1-mini. I've found the 4.1 series to be a ton more reliable, and faster, compared to gpt-5, so that's why I chose it.

Since this relies on OpenAI, you can also swap in any OpenAI-compatible API as well. This means you can swap in [Anthropic], [Gemini], or even open-source LLMs served by vLLM or Ollama.

By modifying the config at ~/.semtools_config.json, you can set the params:

{
  "ask": {
    "api_key": "fake",
    "base_url": "http://localhost:11434/v1",
    "model": "gpt-oss:20b",
    "max_iterations": 10,
    "api_mode": "chat",
  }
}

There's support for both "api_mode": "chat" (which uses chat completions) and "api_mode": "responses" (which uses OpenAI's newer responses API).

Future Work

There's a million things that could be added to this, but at this point I am very open to feedback. Should there be more tools? A dedicated planning step? More config? The sky is the limit honestly. Try it out and see what you think!