Back to Home

Guides & Documentation

Everything you need to know about prompt design, model selections, search systems, and token credits in one place.

Prompt Engineering: Crafting Perfect System Instructions

Prompt engineering is the art of structuring your instructions so that large language models (LLMs) return the exact output format, tone, and logic you require.

Core Best Practices:

  • Define Clear Roles: Start your system instructions with a distinct persona. E.g., "You are an expert senior code reviewer specializing in React performance optimizations."
  • Provide Input Examples (Few-Shot Prompting): Models learn incredibly fast from patterns. Show them 1 or 2 examples of input-to-output pairs before asking the main question.
  • Enforce Negative Constraints: Specify what the model should not do. E.g., "Do not include explanations outside the code block" or "Never assume user credentials."
  • Structured Output Formats: Request outputs in standard structures like Markdown headings, lists, or raw JSON blocks to make parsing easier.
# Example Persona Structure:
- Role: Senior Technical Documenter
- Tone: Professional, structured, clear
- Constraints: Avoid passive voice, use bullet points, maximum 3 paragraphs.

Model Guide: Gemini 1.5 Pro vs Claude 3.5 Sonnet

52bangla AI integrates industry-leading models optimized for various weights. Choose the model that matches your task complexity and token balance:

Model NameStrong SuitToken CostBest For
52bangla AISpeed, Bengali languageFreeDaily tasks, quick translations, general chat
Gemini 1.5 FlashMassive context window, filesFreeAnalyzing large files, long articles, multi-modal inputs
Gemini 1.5 ProLogical reasoning, math, research2x tokensComplex analytical math, structural code writing, deep research
Claude 3.5 SonnetCode syntax, creative prose5x tokensSoftware development, debugging, highly sophisticated copywriting

Understanding Search: Web Search vs Deep Search

Need to fetch facts from the live internet? We offer two unique search modes to keep the AI updated with live statistics and websites:

1. Web Search (Standard Mode)

Performs a quick search across search index engines to pull direct results (up to 5 links) to clarify facts, code docs, or news. Best for quick lookup queries (e.g. "What are today's exchange rates?").

2. Deep Search (Research Mode)

Uses advanced neural indexing (powered by the Exa API) to query the web comprehensively. It pulls up to 10 top results, parses broad text structures, correlates multiple articles, and runs high-context reasoning.

This is a Premium feature exclusively available to Pro and Enterprise tiers, perfect for academic writing, market analysis, and comparative coding research.

Token Cost Efficiency: Getting the Most Out of Credits

Each query you send consumes tokens based on model weights. To make sure you don't drain your balance unnecessarily, follow these tips:

  • Batch Your Questions: Instead of sending multiple short messages (which repeatedly sends the whole chat history, consuming double inputs), ask related questions in a single structured prompt.
  • Clear Inactive Chat Threads: Start a new conversation thread when switching topics. LLMs read the entire current thread history as context; a longer thread consumes significantly more tokens per message.
  • Use Free Models for Drafts: Do initial translation or basic writing using 52bangla AI or Gemini 1.5 Flash. Switch to Claude 3.5 Sonnet only for formatting final adjustments or complex code reviews.