rag-query(1)
==========

NAME
----
rag-query - Query on knowledge-bases

SYNOPSIS
--------
[verse]
'rag query' <query> [<options>]

'rag query --interactive | -i | --multi-turn' [<options>]

'rag query --agent' <query> [<options>]

DESCRIPTION
-----------
Ask AI about the knowledge-base.

There're 3 modes: normal, `--interactive` and `--agent`.

In normal mode, it runs a normal pipeline. It's the simplest way to get an
answer, but still powerful. Run `rag help pipeline` to learn more.

In `--interactive` mode, it'll open an interactive shell in a terminal. You
can have conversations with the model in the terminal. In the interative mode,
you have to press Ctrl+D to enter an input.

In `--agent` mode, an agent browses files and tries to answer your question.
It's the strongest and the most expensive way to get an answer. You might want
to run `rag gc --audit` before running an agent and run `rag audit` to see how
much it costs. It's an overkill to use an agent for a simple question.

OPTIONS
-------
--model <model>::
        An LLM model to query. If it's not set, it's default to `config.model`.

--schema <schema>::
        `--schema` allows you to use pdl schemas with this command. Run
        `rag help pdl-format` to learn more about pdl schemas. It helps the
        model generate an output that follows this schema. For example,
        `rag query <query> --schema="[int]"` will dump an array of integers to
        stdout. You can pipe that array to another program. If `--schema` is
        enabled and the schema is json, like `[int]` or
        `{ name: str, age: int }`, you can always pipe stdout to a json
        processor.

        Please note that it *helps* the model follow the schema, not *forces*
        it. You have to write a nice prompt that asks it to follow the schema.
        The output is largely dependent on the prompt, and `--schema` can only
        correct small mistakes.

        If the model isn't smart enough or the prompt is not clear, it may fail
        to generate a response with valid schema. In that case, it'll dump a
        string "null" to stdout, which is still a valid json. You always have
        to check the output because LLMs can always fail.

        You cannot use `--schema` option in `--interactive` mode.

--continue <uid>::
        You can continue previous conversation. If it's `--interactive` mode,
        it loads the history. Otherwise, it asks a follow up question. You
        cannot use it in agent mode, but it's coming soon.

        You can find query history uids with `ls-queries` command.

--json::
        If `--json` is set, it dumps the result as a json. The json contains
        AI's response and retrieved chunks. `--json` option and `--schema`
        option are very different. Unlike `--schema`, this option does not
        affect how LLM generates responses. In most cases, you don't need
        `--json` when `--schema` is enabled.

--max-summaries <n>::
        It overrides `max_summaries` config. It's temporary and it doesn't
        write to the config files. Run `rag help config-reference` to learn
        more about it.

--max-retrieval <n>::
        It overrides `max_retrieval` config. It's temporary and it doesn't
        write to the config files. Run `rag help config-reference` to learn
        more about it.

--enable-ii | --disable-ii::
        It overrides `enable_ii` config. It's temporary and it doesn't
        write to the config files. Run `rag help config-reference` to learn
        more about it.

--enable-rag | --disable-rag::
        It overrides `enable_rag` config. It's temporary and it doesn't
        write to the config files. Run `rag help config-reference` to learn
        more about it.

--[no-]super-rerank::
        It overrides `super_rerank` config. It's temporary and it doesn't
        write to the config files. Run `rag help config-reference` to learn
        more about it.