sanjaya 0.2.0

Extensible RLM agent framework with video understanding

Sanjaya

Sanjaya is a Python library for building evidence-first RLM (Recursive Language Model) agents.

It runs a loop where the model writes Python, executes it in a sandboxed REPL, inspects results, calls tools, and iterates until it can return a grounded answer.

Why the name?

In the Mahabharata, Sanjaya narrates events to a blind king with "divine sight." This project follows the same idea: observe evidence and report what happened.

Install

pip install sanjaya

With extras:

pip install "sanjaya[video]"      # video analysis
pip install "sanjaya[image]"      # image analysis (Pillow, HEIC, SVG)
pip install "sanjaya[document]"   # PDF/EPUB/PPTX/Markdown/text parsing
pip install "sanjaya[tracing]"    # tracing integrations
pip install "sanjaya[all]"        # everything

Configuration

Set at least:

• OPENROUTER_API_KEY

Optional (depends on your model/provider setup):

• OPENAI_API_KEY
• MOONDREAM_API_KEY
• LOGFIRE_TOKEN

Quick start

from sanjaya import Agent

agent = Agent(max_iterations=8)

answer = agent.ask("What is the capital of France?")
print(answer.text)

Video analysis

from sanjaya import Agent

agent = Agent(max_iterations=12)

answer = agent.ask(
    "How many people are speaking and what are they discussing?",
    video="/path/to/video.mp4",
    # subtitle="/path/to/subtitle.json",  # optional
)

print(answer.text)
print(answer.evidence)

When video=... is provided, VideoToolkit is auto-registered if you have not already added one.

Document analysis

from sanjaya import Agent

agent = Agent(max_iterations=12)

answer = agent.ask(
    "Summarize the key claims and supporting evidence.",
    document=["/path/to/report.pdf", "/path/to/appendix.md"],
)

print(answer.text)
print(answer.evidence)

Supported document types include .pdf, .epub, .pptx/.ppt, .md, and .txt.

When document=... is provided, DocumentToolkit is auto-registered if needed.

Image analysis

from sanjaya import Agent

agent = Agent(max_iterations=10)

answer = agent.ask(
    "What text is visible in this screenshot?",
    image="/path/to/screenshot.png",
)

print(answer.text)
print(answer.evidence)

Multiple images:

answer = agent.ask(
    "Compare these two charts and summarize differences.",
    image=["/path/to/chart_q1.png", "/path/to/chart_q2.png"],
)

Supported formats: JPEG, PNG, WebP, GIF, TIFF, BMP, HEIC (requires sanjaya[image]), SVG (requires sanjaya[image]).

When image=... is provided, ImageToolkit is auto-registered if needed.

Custom tools

from sanjaya import Agent, Toolkit, tool

class MyToolkit(Toolkit):
    @tool
    def lookup(self, query: str) -> str:
        return f"result for: {query}"

agent = Agent().use(MyToolkit())
answer = agent.ask("Find the latest status update")
print(answer.text)

Answer object

Agent.ask() returns an Answer object with:

• text
• evidence
• iterations
• input_tokens, output_tokens, cost_usd
• wall_time_s

Development (repo)

If you are working from the repository:

• just dev — API + UI
• just api — FastAPI backend (:8000)
• just ui — Next.js frontend (:5100)
• just demo — run demo prompts

Status

Sanjaya is actively developed. Core video + document flows are working, and evaluation/dashboard tooling is still evolving.