PORF

Publish Original Research Fast

PORF is a minimalist Python library that writes research articles on any topic. It runs a multi-expert roundtable discussion, collects sources, and produces a polished long-form article with citations.

Named after Porfiry Petrovich — the literary AI from Victor Pelevin's iPhuck 10.

Installation

pip install porf

# With interactive CLI (recommended)
pip install porf[cli]

# With DuckDuckGo search (free, no API key)
pip install porf[search]

# With semantic retrieval (better source routing)
pip install porf[semantic]

# Everything
pip install porf[all]

Quick Start

CLI

# Interactive wizard
porf

# One-liner
porf "Impact of AI on creative industries" -p deep -t popular -o report.md

# Custom article length
porf "AI in healthcare" -p balanced -w 5000 -o report.md

# Multi-language search, output in Russian
porf "AI in healthcare" -l en,ru --out-lang ru -o report.md

Python API

from porf import research

report = research(
    topic="The Impact of Large Language Models on Software Development",
    profile="balanced",
    style="analytical",
    target_words=5000,
)

print(report.markdown)

How It Works

PORF follows a four-stage pipeline:

1. EXPLORE
   Decompose topic → search queries → collect initial findings

2. BOOTSTRAP
   Generate expert panel with clashing perspectives
   Design article outline (sections with claims, optional subsections)

3. ROUNDTABLE (multi-round)
   Each expert researches their assigned section:
     generate queries → search → draft with citations
   Information routed to mind map via embeddings
   Sections auto-lock when they converge
   Sections can split into subsections during research

4. WRITE
   Narrative plan → write body sections (sequential) →
   write intro → write conclusion → assemble with hierarchy

The roundtable produces research drafts with citations. The writer transforms them into a polished article with narrative flow, style-appropriate prose, and consistent citation numbering.

Profiles & Styles

Profiles control research depth and article length:

Profile	Experts	Max rounds	Lock threshold	Target words
quick	3	10	5 sources	~2000
balanced	4	15	8 sources	~4000
deep	5	20	12 sources	~6000

Styles control article tone and writing techniques:

Style	Tone
analytical	Systematic, fair, data-driven
academic	Precise, authoritative, theory-friendly
journalistic	Direct, concrete, fact-first
popular	Conversational, accessible, theory-light
essay	Reflective, personal, exploratory

porf "Quantum computing" -p deep -t journalistic

CLI Reference

porf [TOPIC] [OPTIONS]

Arguments:
  TOPIC                    Research topic (wizard mode if omitted)

Options:
  -p, --profile PROFILE    quick, balanced, deep (default: balanced)
  -t, --style STYLE        analytical, academic, journalistic, popular, essay
  -w, --words N            Target article length in words (overrides profile)
  -m, --model MODEL        LLM model (default: anthropic/claude-sonnet-4-20250514)
  -s, --search ENGINE      duckduckgo, tavily, brave, serper, searxng
  -l, --lang LANG          Search language(s): auto, en, ru, or en,ru
  --out-lang LANG          Output language: auto, en, ru, etc.
  -o, --output FILE        Output file path
  --api-base URL           API base URL for local models

Search Engines

Engine	API Key	Notes
duckduckgo	Not needed	Default, free
tavily	TAVILY_API_KEY	Best for research
brave	BRAVE_API_KEY	Good quality
serper	SERPER_API_KEY	Google results
searxng	Not needed	Self-hosted

Local Models

from porf import research

# LM Studio
report = research(
    topic="...",
    profile="quick",
    model="openai/local-model",
    api_base="http://localhost:1234/v1",
)

# Ollama
report = research(
    topic="...",
    model="ollama/llama3",
    api_base="http://localhost:11434",
)

Output

report.markdown   # Full article with citations
report.sections   # List of Section objects
report.sources    # List of Source objects
report.to_dict()  # JSON-serializable dict

License

Apache 2.0