Forem: Suraj Bera

Day 4: ReAct - Reasoning + Acting upon(Prompting Technique)

Suraj Bera — Mon, 04 May 2026 15:04:12 +0000

Big picture of ReAct:

AI doesn't just answer immediately, it goes through stages.
think -> plan -> action -> observe -> output(total 5 steps)

ReAct = Forcing the AI to show its work step by step before arriving at a conclusion, just like how a student arrives to the solution of a math problem step by step.

Here is the simplest mental model for ReAct:

        ┌──────────────────────────────┐
        │    YOU send a question       │
        └──────────────┬───────────────┘
                       ▼
        ┌──────────────────────────────┐
        │  AI replies with JSON like   │
        │  { "step": "...", ... }      │
        └──────────────┬───────────────┘
                       ▼
        ┌──────────────────────────────┐
        │ Python looks at "step" and:  │
        │  • plan  → loop again        │
        │  • action → run a tool       |
        │  • output → print & STOP     │
        └──────────────────────────────┘

Comparing Chain of Thought(another prompting technique) vs ReAct:
CoT - In this technique, the LLM is instructed to generate its intermediate reasoning steps as part of the output, instead of jumping straight to the final answer.
ReAct - CoT only thinks, ReAct thinks and does stuffs.

Here is the code

Links to Day 3 and Day 2

Day 3: Prompting Techniques in AI (Part 1)

Suraj Bera — Mon, 04 May 2026 03:15:05 +0000

AI Prompting techniques: Zero-shot, One-shot, Few-shot

After using the ChatGPT and other AI tools, I used to think prompts were just simple text inputs that AI models magically processed. But as mentioned in my Day 1 post: AI models are just next-word predictors, not thinkers. They predict based on training data (though modern ones now use real-time search and tool calling for better results).

The Basics
Simplest python code snippet for getting response from AI

response = client.responses.create(
    model="gpt-5-nano",
    input=[
        {"role": "system", "content": SYSTEM_PROMPT},
        {"role": "user", "content": "How do I reverse a list in Python?"},
    ]
)

The most important point to note: role key has 3 values: system, user, assistant. The system prompt sets the behaviour.

1. Zero-shot prompting
No examples -- just instructions in the system prompt

SYSTEM_PROMPT = """
You are a helpful assistant that only answers Python programming questions.
If the user asks about anything else, politely decline.
"""

2. One-shot prompting
Provide one example how model should respond.

input=[
    {"role": "system", "content": SYSTEM_PROMPT},
    {"role": "user", "content": "What is the capital of France?"},
    {"role": "assistant", "content": "I'm sorry, I can only help with Python programming questions."},
    {"role": "user", "content": "How do I reverse a list in Python?"},
]

3. Few-shot prompting
Provide multiple examples so the model responds more precisely:

input=[
    {"role": "system", "content": SYSTEM_PROMPT},
    {"role": "user", "content": "How to code a binary tree in Python?"},
    {"role": "assistant", "content": "Sure, here's an implementation..."},
    {"role": "user", "content": "What is the weather today?"},
    {"role": "assistant", "content": "I'm sorry, I can only help with Python programming questions."},
    {"role": "user", "content": "What is the capital of France?"},
    {"role": "assistant", "content": "I'm sorry, I can only help with Python programming questions."},
    {"role": "user", "content": "Why is 75% attendance required for the exam?"},
]

So as a beginner trying to get into AI engineering, you need to shift your mindset from just chatting with AI mindlessly to designing system prompt for your own AI product.

Complete code for few-shot-prompting

Links to day 1 and day 2

Part 2: Vector Embeddings in simplest terms

Suraj Bera — Sun, 03 May 2026 05:03:53 +0000

This is my Day 2 of learning AI fundamentals where I will be covering the following concepts:

Vector Embeddings
How Tokenisation and Vector Embeddings relate to each other

Vector embeddings:

Vector embeddings is the process of turning each token id(generated during tokenisation) into high dimensional vector where semantic similarity results into geometric closeness. Think of it like this: dog is closer to puppy, also closer to dog food. But dog is not closer to car or petrol.

When we use embeddings?

Recommendations: Suggest similar songs, videos, movies, products
Search: Get search results when keywords don't match
Cluster: Grouping related concepts together.

A beginner might be confused in terms like: Vector, High Dimensional.

This is an example of a vector: [0.9, 0.8, 0.1]. Array/List/Vector all mean the same thing. 'List' is just a plain english, 'array' is the programming term, vector is the math/ml term.
High Dimensional: Multi-dimensional just means more than 1 - could be 2D, 3D, 10D,... too vague. But high dimensional specifically means hundreds of thousands of dimensions(Open AI's text-embedding-3-small = 1536 dimensions)

There are 2 types of Search:

Lexical Search: Exact words/characters search
Semantic Search: Meaning/Intent search

Vector embedding enables semantic search

How tokenisation & vector embeddings are connected together?

text-tokenisation --> token Ids-embedding lookup --> vectors --> transformers
"hello" --> [221728] --> [0.21,-0.44,...]

Vector Embedding Visualizer

Here is the link to part 1