Great video and I came to the same conclusion. I use a inexpensive model to do tool calls (and perplexity searches) and then feed a perfectly formated query to a reasoning model. Then use a local (or gemini flash) to format it.
Why is nobody ever mentioning in any RUclips video at all since this thing first came out that you can now use R1 AND Websearch at the same time? That's huge and everybody is literally freaking blind about it lol
It maybe censure, all person thing chatgtp, its god, but nothing or few person, speak claude. Or r1 deepseek, this people pay 200 dollar its normal, they think i pay 200$ i need say this is the best, and i rescue my inversiones. If see task openai its. Claude maybe existe last years and nothing say... This stage ios VS android. In other video deepseek big model nivel exact with geminis, and level up chatgtp. But i think chatgtp this is no real agi , because literal base culture US. But not word becuase if you ask for Oeste, for real history american , but this not exist. It false.
I've been trying to build non agentic script to do tavily research with llm processes to refine search, clean results, use results, structure and build reports, and keep getting bogged down in multiple calls and handling data models. Just realized i can just use a pydantic AI agent simply as a tool in that process to run through those without me handling the overhead and still get the structured output for chaining processes i need! If this works it will be amazing....
Out of curiosity, for json output, why not use one of the many markdown converter packages in python? Seems like R1 already produces pretty good markdown. Would that give more consistent results than making another LLM call? Thanks for a great video.
Great video, thanks! I'll be honest, I'm afraid of putting more effort on these kind of tricks, it might go absolute in a few months or weeks! The racing is getting intense, we'll see what o3, Opus (according to some leaks)... will bring to the table.
Not sure about the JSON output, but you can make it output XML (and wrap the answers into CDATA sections inside the tags) by having an appropriate prompt. At least this works with the local distilled versions. But thanks for a good idea: I could try getting the output from the r1:14b and feed it into e.g. qwen2.5-coder:3b as a formatting agent.
Apparently with reasoning models in the system prompt you aren’t supposed to give the model the whole persona thing such as “you are a reasoning model…”
This recommendation is specific to R1 and based on the observation that it performs worse in some benchmarks when given a system prompt than without it. Which only goes to show how stupid these models still are.
Great video and I came to the same conclusion. I use a inexpensive model to do tool calls (and perplexity searches) and then feed a perfectly formated query to a reasoning model. Then use a local (or gemini flash) to format it.
This is also what I am doing. Reasoning model for intellect, non reasoning for structured output and tool calling.
combining with gemini flash to clean the output is genius move
Why is nobody ever mentioning in any RUclips video at all since this thing first came out that you can now use R1 AND Websearch at the same time? That's huge and everybody is literally freaking blind about it lol
Exactly
It maybe censure, all person thing chatgtp, its god, but nothing or few person, speak claude. Or r1 deepseek, this people pay 200 dollar its normal, they think i pay 200$ i need say this is the best, and i rescue my inversiones. If see task openai its. Claude maybe existe last years and nothing say...
This stage ios VS android. In other video deepseek big model nivel exact with geminis, and level up chatgtp.
But i think chatgtp this is no real agi , because literal base culture US. But not word becuase if you ask for Oeste, for real history american , but this not exist. It false.
@@ajwo5984 In open router chat it worked from day one lol
good stuff! I never thought of using the reasoning model as a tool!
I've been trying to build non agentic script to do tavily research with llm processes to refine search, clean results, use results, structure and build reports, and keep getting bogged down in multiple calls and handling data models. Just realized i can just use a pydantic AI agent simply as a tool in that process to run through those without me handling the overhead and still get the structured output for chaining processes i need! If this works it will be amazing....
Thank you so much for providing Hindi Track!
Out of curiosity, for json output, why not use one of the many markdown converter packages in python? Seems like R1 already produces pretty good markdown. Would that give more consistent results than making another LLM call? Thanks for a great video.
Great video, thanks! I'll be honest, I'm afraid of putting more effort on these kind of tricks, it might go absolute in a few months or weeks! The racing is getting intense, we'll see what o3, Opus (according to some leaks)... will bring to the table.
Not sure about the JSON output, but you can make it output XML (and wrap the answers into CDATA sections inside the tags) by having an appropriate prompt. At least this works with the local distilled versions.
But thanks for a good idea: I could try getting the output from the r1:14b and feed it into e.g. qwen2.5-coder:3b as a formatting agent.
I am actually building an mcp server with reasoning tools and R1 as the sole contender all for Roo Cline.
Precisely what I need
Nice video, can you try Kimi 1.5? They ve just launched the new model it's also open source
Thanks 🎉🎉🎉
Apparently with reasoning models in the system prompt you aren’t supposed to give the model the whole persona thing such as “you are a reasoning model…”
What should you put there instead, instructions?
yeah just more nuance, it’s implicit that it will think step by step w reasoning models. not like it would hurt to say that tho, it’s just redundant
This recommendation is specific to R1 and based on the observation that it performs worse in some benchmarks when given a system prompt than without it. Which only goes to show how stupid these models still are.
Great video
I believe and after consulting with DeepSeek that smolagents being that it outputs write its own code in Python. there is not a Json structure needed.
Please make a video about how to use distill models like DeepSeek-R1-Distill-Qwen-32B. thanks
Anything particularly you want to use it for ?
@@samwitteveenai for document question answering, recommender sys.
and can we use this model as agentic usecase like using in langchain?
using it to "weaponize" some cybersecurity tools, for penetration testing work.
dude we need examples
Love it
Best Part our Politicians are still fighting over Tic Tok. 🤣🤣🤣
Actually TikTok has more relevance to the power structure of the world than "reasoning" models that still trip up on their own shoelaces.
Cool
poor thing gemini 1.5 : you never give your own opinion! hahaha
lol it’s not allowed to have an opinion now 2.0 is here
Knowing how these models work, it will give you its opinion anyway (from time to time, unpredictably).
ollama locally is the cheapest way to do it, dont use apis that you have to pay forunless you really have to!