DeepSeek R1 for Structured Agents



AI Summary

Summary of Video Transcript

  • The video discusses how to use the Deep Seek reasoning model for creating structured agents.
  • Deep Seek has added a search feature on their website, allowing simultaneous use of Deep Think and Search.
  • The model currently lacks support for function calling and JSON output, presenting challenges for agent development.
  • The video provides hacks to work around these limitations using Pantic AI.
  • Gemini 2.0 thinking models have similar issues but are expected to improve over time.
  • The video demonstrates using prompt engineering and a prompting framework to get structured responses.
  • The API provides separate thinking and answer tokens, which can be manipulated.
  • The video includes a tutorial on using Google Colab and Pantic AI to set up structured outputs.
  • The API setup is similar to OpenAI’s, with the base URL redefined to api.deep.com.
  • Deep Seek offers a chat model (Deep Seek V3) and a reasoning model (Deep Seek R1).
  • The video shows how to use the Deep Seek API with the Pantic AI setup to get structured outputs.
  • It discusses the use of a cheaper model like Gemini 1.5 Flash for formatting outputs.
  • The video explains how to extract reasoning content and normal content from the model’s output.
  • It also covers using the Reasoner model as a tool within an orchestrated agent setup.
  • The video concludes with the potential of reasoning models for agent applications and plans for future content.

Detailed Instructions and Tips (if present in the transcript)

  • Turn on Deep Think and Search on the Deep Seek website to use both features.
  • Use Pantic AI for structured outputs and easy model swapping.
  • Redefine the base URL for the OpenAI API to api.deep.com for Deep Seek API.
  • Set up Deep Seek chat and reasoning models with API keys.
  • Use the platform at platform.deepseek.com to access the API (no exact URL provided).
  • Use the reasoning content and normal content from the API for structured outputs.
  • Combine reasoning and answer content with formatting tags for structured results.
  • Use a cheaper model like Gemini 1.5 Flash for formatting if needed.
  • Consider using a simpler model for keyword generation rather than the reasoning engine.

(Note: No specific CLI commands, website URLs, or self-promotion were included in the provided text.)