AI Query Assistant

rfx ask lets you query your codebase using natural language. An LLM turns your question into rfx query commands, Reflex runs them against your local index, and (optionally) the LLM answers from the results.

Setup

rfx ask needs an index (rfx index) and an LLM provider. Configure the provider with:

rfx llm config

This interactive wizard sets up your provider, API key, and model. Configuration is stored in ~/.reflex/config.toml and is shared with Pulse. (rfx ask --configure still works, but is deprecated in favor of rfx llm config.)

Check your current setup:

rfx llm status

Provider: anthropic
Model:    claude-3-5-haiku-20241022
API key:  configured (sk-ant-a...****)

Supported providers

Provider	provider value	Default model
OpenAI	openai (default)	gpt-4o-mini
Anthropic	anthropic	claude-3-5-haiku-20241022
OpenRouter	openrouter	anthropic/claude-sonnet-4
OpenAI-compatible	openai-compatible	none (you must set a model and base URL)

The OpenAI-compatible provider works with any server that exposes an OpenAI-style /chat/completions endpoint, such as a local model server. The API key is optional for this provider. Without one, requests are sent with no auth header.

Manual configuration

You can also edit ~/.reflex/config.toml directly:

[semantic]
provider = "anthropic"
# model = "claude-3-5-haiku-20241022"   # optional; overrides the per-provider model below

[credentials]
anthropic_api_key = "sk-ant-..."
anthropic_model = "claude-3-5-haiku-20241022"

The [credentials] section takes <provider>_api_key and <provider>_model for openai, anthropic, openrouter, and openai_compatible. OpenRouter also accepts openrouter_sort (price, speed, or throughput). For an OpenAI-compatible server:

[semantic]
provider = "openai-compatible"

[credentials]
openai_compatible_base_url = "http://localhost:1234/v1"
openai_compatible_model = "qwen2.5-coder"
# openai_compatible_api_key = "..."    # optional

Environment variables

For CI or headless use, you can configure everything from the environment:

Variable	Purpose
REFLEX_PROVIDER	Override the provider
REFLEX_MODEL	Override the model
REFLEX_AI_API_KEY	API key for whichever provider is selected
OPENAI_API_KEY, ANTHROPIC_API_KEY, OPENROUTER_API_KEY, OPENAI_COMPATIBLE_API_KEY	Provider-specific API keys
OPENAI_COMPATIBLE_BASE_URL, OPENAI_COMPATIBLE_MODEL	Fallback base URL and model for openai-compatible
REFLEX_LLM_TIMEOUT_SECONDS	LLM request timeout (default: 30)

API keys are looked up in this order: ~/.reflex/config.toml, then REFLEX_AI_API_KEY, then the provider-specific variable.

Interactive mode

Start a conversation about your code:

rfx ask

With no question, rfx ask opens an interactive chat TUI (same as rfx ask -i) with full conversation history. The assistant searches your codebase to ground its answers.

>> How does authentication work in this project?
>> What files would I need to change to add OAuth support?
>> Show me the error handling pattern used in the API handlers.

Slash commands inside the chat:

Command	Action
/clear	Clear conversation history (also Ctrl+L)
/compact	Summarize old messages to save tokens (also Ctrl+K)
/model [provider] [model]	Show or change provider/model
/help	Show help

Ctrl+U clears the input, Up/Down and PgUp/PgDn scroll, and Ctrl+C quits.

One-shot mode

Ask a single question from the command line:

rfx ask "What does the handleRequest function do?"

By default, one-shot mode runs the generated queries right away. It prints them under a Generated Queries: heading, followed by the raw matches grouped by file (Found N total results across M files:).

Modes

Answer mode

Add --answer to get a conversational answer based on the search results, instead of the raw matches:

rfx ask "What database does this project use?" --answer

Agentic mode

Enables multi-step reasoning. The assistant decides whether it needs more context, gathers it with tools, generates queries, runs them, evaluates the results, and refines them if needed:

rfx ask "How would I add a new API endpoint?" --agentic --answer

Flag	Description
--max-iterations <N>	Maximum refinement iterations (default: 2)
--no-eval	Skip result evaluation
--show-reasoning	Show LLM reasoning at each phase
--verbose	Show tool results and details
--quiet	Suppress progress output

Confirm before running

rfx ask "Find all TODO comments" --execute

With --execute, rfx ask shows the generated queries and asks Run these queries against your codebase? [y/N] before running them. Anything but y aborts.

Caution

The v2.0.0 --help text describes --execute as “Execute queries immediately without confirmation”. The code does the opposite: a plain one-shot rfx ask already runs queries immediately, and --execute adds the confirmation prompt.

JSON output

rfx ask "Find all TODO comments" --json --pretty

Returns the generated queries, the results, the total count, and (with --answer) the answer as JSON, without progress output.

Provider selection

Override the configured provider for a single query:

rfx ask "Explain the auth flow" --provider anthropic
rfx ask "Explain the auth flow" --provider openai-compatible

Project context

If your project has a REFLEX.md file in the root, rfx ask automatically includes it as context. This helps the assistant understand your project’s architecture, conventions, and terminology. See Configuration for details.

You can also inject extra context for a single question, for example from rfx context:

rfx ask "find auth" --additional-context "$(rfx context --framework)"

Use --debug to print the full LLM prompts.

Next steps

• AI Integration — MCP server and JSON output for agent workflows
• Configuration — project configuration
• CLI Commands — full rfx ask reference