LLM - Day 3 - Prompt Eng 2

REACT Framework ( Reasoning & Action )

• LLM is model or a model file.
• இப்ப ஒரு LLM Model வந்து என்கிட்ட இருக்கு , அந்த Model வந்து ஏதோ ஒரு set of dataவால train ஆகியிருக்கும் நான் இப்ப query கொடுத்தா அது எனக்கு ஒரு response தரும்.
• இப்ப நான் அது கிட்ட போயிட்டு இன்னைக்கு gold rate எவ்ளோ என்று கேட்டால் அது தருமா ? தராதா ?
• LLM --> Algorithm and pre-trained data is a Model. Also LLM is a file.

• If LLM can't do anything , it will expect an ACTION or some other tools.
• Tools are called as programs.
• Programs are called scripts.
• LLM + Tools ( Tools nothing but a script ).
• Can a LLM call another file ? Yes , you can call the tool is AGENT.
• Its called TOOL CALLING --> He is nothing but AGENT.

In framework we have RICE Pattern.

• For this pattern we give --> Role , Input , Contraints , Expectations will be provided and ask the query.

In framework we have CRISP Pattern.

• For this pattern we give --> Context , Role , Instruction , Style , Purpose.
• Eg., Blog writing

TOOL CALLING - AGENT

• run_example.py file will import all items from the techniques folder and then running each one functionalities.
• need to be installed all these to run. pip install -r file_name , requirement file.

• To create a Agent " langchanin " is enough.
• LLM + Tool --> to orchestrate these two is AGENT.
• .env will have API key of the groq. ( which will connect to the cloud LLM ) why we are connecting like this ? Because we can't run the LLM in the local as we are going to use gpt-oss-120b.
• Render the messages is also available.

Zero-Shot

• run_prompt
• **ChatPromptTemplate **is from langchain package.
• All will be passed in LIST.
• No example will be given for this Zero-shot

Few-Shot

• Example will be given.

Tree of thought

• Option based [TBD]

Chain of thought
Role based
Instruction based
Contextual Prompting
Self consistency
React

Notes :

• Many frameworks are there --> like CARS.
• Groq --> Infra provider for LLM.
• llama3 & gpt-oss 120B --> these are two models which are predominantly used.
• langchain is more stable to create a agent.
• Temperature --> based on facts or closely/choice.
• Don't memorize the code , because each framework is differemt. Just get the flow.
• How the model are choosen ? Hugging face ( they will give the model ) , ollama.
• Image --> Blip Model.
• Text --> Qwen Model.