Skip to content

Usage Guide

This page covers all CLI commands and common workflows.


Quick Reference

Command Purpose
ask "question" Research question → auto-discover papers → full pipeline
ask "question" --session ID Continue an existing session
papers "query" Interactive paper explorer — search, select, ask
paper <arxiv_id> "question" Deep-dive into a specific paper
search "query" Search ArXiv (listing only, no analysis)
sessions List all conversation sessions
workflow show Display current pipeline configuration
workflow toggle <stage> <on\|off> Enable/disable a workflow stage
workflow reorder <s1,s2,...> Reorder workflow stages
artifacts list List stored reasoning artifacts

Command Details

ask — Research Questions

Ask a research question and let the system discover relevant papers:

# Basic usage
python -m src.main ask "What are the latest advances in neural architecture search?"

# Continue a previous conversation
python -m src.main ask "How does this compare to random search?" --session abc123

What happens: 1. Question is triaged (general/conceptual/research) 2. If research question: ArXiv papers discovered and downloaded 3. Papers chunked, indexed, and retrieved 4. Context compressed via ScaleDown 5. Full anti-hallucination pipeline runs (COT → Verify → Critique) 6. Answer displayed with citations 7. Session saved for follow-up questions

Expected time: - General questions: ~5s (direct answer) - Research questions: ~30-60s (paper discovery + full pipeline)


papers — Interactive Paper Explorer

Search, browse, and analyze papers interactively:

python -m src.main papers "attention mechanism transformers"

Interactive Commands: - Type text: Ask a question about the currently selected paper - Type a number: Switch to a different paper from the list - back: Return to the paper list - list: Show the paper list again - s: Start a new search - q: Quit the explorer

Features: - Persistent session: All interactions share the same session - No refetching: Papers are cached — follow-up questions are instant - Conversation history: LLM sees previous Q&A for better context - Full pipeline: Each answer goes through COT → Verify

Example session:

> papers "attention mechanism transformers"
[Paper list displayed]

Select a paper (1-10), or type 's' to search again, 'q' to quit: 1
[Paper #1 selected: "Attention Is All You Need"]

Ask a question (or 'back' to list, 'q' to quit): What is the multi-head attention mechanism?
[Full pipeline runs → Answer displayed with citations]

Ask a question (or 'back' to list, 'q' to quit): How does it differ from single-head attention?
[Follow-up answered using session history]

Ask a question (or 'back' to list, 'q' to quit): back
[Returns to paper list]

Select a paper (1-10), or type 's' to search again, 'q' to quit: q

paper — Deep-Dive into a Specific Paper

Analyze a known ArXiv paper:

python -m src.main paper 1706.03762 "What is the multi-head attention mechanism?"

What happens: 1. Downloads ArXiv paper 1706.03762 (if not cached) 2. Extracts text and indexes it 3. Runs full pipeline with paper-specific grounding 4. Answer focused ONLY on content from this paper

Paper-specific grounding means: - System prompt explicitly says "analyze THIS specific paper" - LLM instructed NOT to use training data - Must cite specific sections, equations, figures, or tables - If paper doesn't mention something, it must say so


Search ArXiv without analysis:

python -m src.main search "graph neural networks"

Output: - Top 10 ArXiv results with titles, authors, and abstracts - No paper download or analysis - Useful for quickly finding relevant papers


sessions — View Conversation History

List all saved sessions:

python -m src.main sessions

Output: - Session IDs - Creation timestamps - Number of Q&A turns - Papers involved

Use a session ID with ask --session to continue a conversation.


workflow — Configure Pipeline

View and modify the reasoning pipeline:

Show Current Configuration

python -m src.main workflow show

Output: - List of stages in order - Enabled/disabled status for each

Toggle Stages

# Enable self-critique
python -m src.main workflow toggle self_critique on

# Disable self-verification (not recommended!)
python -m src.main workflow toggle self_verify off

Reorder Stages

# Standard order
python -m src.main workflow reorder cot,self_verify,self_critique

# Custom order (e.g., critique before verify)
python -m src.main workflow reorder cot,self_critique,self_verify

artifacts — View Stored Outputs

List all stored reasoning artifacts:

python -m src.main artifacts list

Output: - Artifact IDs - Stage (cot, self_verify, self_critique) - Timestamps - File sizes

Artifacts are stored in artifacts/ as compressed markdown files.


Workflow Examples

Example 1: Quick Research

# Single command gets you a cited answer
python -m src.main ask "What are transformers in NLP?"

Result: - ArXiv papers discovered - Relevant chunks retrieved and compressed - COT reasoning with citations - Verification of all citations - Critique of answer quality - Final answer displayed in terminal


Example 2: Multi-Turn Conversation

# First question
python -m src.main ask "What is neural architecture search?"
# Note the session ID in the output: abc123

# Follow-up
python -m src.main ask "What are the main challenges?" --session abc123

# Another follow-up
python -m src.main ask "How does DARTS address these?" --session abc123

Result: - Each follow-up has context from previous turns - Papers are cached (no re-downloading) - Session history grows with each interaction


Example 3: Interactive Paper Exploration

python -m src.main papers "transformers"

# Select paper #1
1

# Ask questions
What is the encoder structure?
How many layers does it have?
What is positional encoding?

# Switch to paper #3
3

# Compare
How does this differ from the original transformer?

# Exit
q

Result: - Seamless switching between papers - All questions share one session - Full pipeline runs for each answer


Example 4: Custom Workflow

# Disable critique for faster responses
python -m src.main workflow toggle self_critique off

# Ask question (skips critique stage)
python -m src.main ask "What is gradient descent?"

# Re-enable critique
python -m src.main workflow toggle self_critique on

Result: - Workflow configuration persists across runs - Faster responses when critique is disabled - Flexibility to trade quality for speed


Next: Methodology

See Methodology to understand the technical approach.