On June 5, 2026, the residency focused on Prompt Engineering—the science of writing clear instructions to align AI outputs. We studied the RTACF (Role, Task, Audience, Context, Format) framework. Role prompting narrows down the search space in the model's parameters, ensuring appropriate tone and vocabulary. We compared unstructured prompts to highly structured ones, demonstrating how specificity drives output quality.

1. Prompt Engineering: Directing Generative Logic

We focused on role prompting and structured instructions. Organising queries using Role, Task, Audience, Context, and Format templates restricts the model's statistical parameters, shifting outputs from generic wiki descriptions to targeted execution payloads.

2. LLM Hyperparameters: Temperature and Token Mechanics

We analyzed core inference parameters: Temperature and Tokens. Temperature controls the randomness of next-token selection: low temperature (e.g., 0.2) results in deterministic, factual, and consistent responses (ideal for code and math), while high temperature (e.g., 1.0) increases creativity. Text is processed as Tokens—sub-word chunks determined by BPE tokenizers. We discussed context window constraints and token-based API pricing models.

3. Advanced Customization: Fine-Tuning & RAG Pipelines

Rashmi introduced advanced AI architecture patterns: Pre-training (costly training from scratch), Fine-Tuning (adapting weights on instruct datasets), and Vector Embeddings (semantic numerical projections). We then explored RAG (Retrieval-Augmented Generation), where document chunks are fetched from vector databases (like Qdrant) via Cosine Similarity to enrich prompt contexts dynamically, eliminating hallucinations and knowledge-cutoff limits.

Key Learnings

  • Constructing structured prompt interfaces using the RTACF framework.
  • Testing softmax temperature coefficients (0.2 vs. 1.0) for facts vs. creative tasks.
  • Diagramming Retrieval-Augmented Generation (RAG) system pipelines.
  • Explaining pre-training, fine-tuning, and semantic vector embedding metrics.

Tools & Stack

  • Gemini API
  • Vector Embeddings
  • Qdrant DB
  • Prompt Engineering
  • Python 3.11

Challenges Overcome

  • Optimizing prompt layouts to prevent context window token overflow.
  • Aligning model vocabulary output using few-shot formatting examples.

Task to be Performed

  • Evaluate prompt performance using unstructured vs. structured role-prompting formats.
  • Perform temperature coefficient parameters testing (0.2 vs. 1.0) on facts vs. creativity tasks.
  • Analyze Byte-Pair Encoding (BPE) token boundaries and API consumption constraints.
  • Create system personality prompts for StudyBot with role contexts and few-shot rules.

Related Logs

Day 00

Onboarding & Exploration: Mapping the Core AI & Python Blueprint

Onboarding at Virtual Height, exploring Python-based AI development pipelines, setting up sandboxes, and defining our core AI & Prompt Engineering curriculum.

May 29, 2026 • 4 min readRead Log ➔
Day 01

Day 01: Foundations of AI, Python Lab, and the Velocity of Domain Mastery

Onboarding under Senior AI Trainer Rashmi, mapping the formal pillars of AI from ML to DL, mastering cross-platform Python execution, and demonstrating custom LLM and reverse engineering portfolios to the cohort.

Jun 01, 2026 • 4 min readRead Log ➔