Data Science & LLM Guide.

Transformers removed sequential processing constraints by introducing self-attention, allowing parallel computation over tokens. This solved long-range dependency issues and significantly improved training efficiency on large-scale datasets.

Self-attention computes Query, Key, and Value matrices for each token. The attention score is calculated via dot-product similarity between Q and K, scaled, and normalized using softmax. This determines how much each token should focus on others in the sequence.

Embeddings map discrete tokens into dense vector spaces where semantic similarity is preserved. This enables LLMs to generalize meaning, perform similarity search, and power retrieval systems like RAG.

A production RAG pipeline includes: document ingestion, chunking strategy, embedding model, vector database (like FAISS or Pinecone), retrieval logic (top-k search), and an LLM generation layer conditioned on retrieved context.

Poor chunking leads to semantic fragmentation or context dilution. Optimal chunking preserves meaning boundaries (paragraphs, sections) and improves retrieval precision while reducing hallucination risk.

Vector databases store high-dimensional embeddings and perform approximate nearest neighbor search (ANN), unlike SQL systems that rely on exact key matching. They are optimized for semantic similarity rather than structured queries.

Hallucinations occur when retrieved context is irrelevant, incomplete, or overridden by the LLM’s internal priors. They can be reduced using better embeddings, reranking models, and stricter prompt grounding.

Fine-tuning is preferred when you need behavior learning (tone, format, classification rules). RAG is better for factual knowledge updates. In practice, many systems combine both.

LoRA (Low-Rank Adaptation) reduces training cost by injecting trainable low-rank matrices into frozen weights, allowing efficient fine-tuning of large models without full retraining.

Reinforcement Learning from Human Feedback aligns model outputs with human preferences by training a reward model and optimizing responses using policy gradients.

Training is compute-heavy and offline, focused on learning parameters. Serving is real-time inference optimized for latency, scalability, and cost efficiency.

Concept drift occurs when real-world data distribution changes over time, making trained models less accurate. Monitoring requires tracking input/output distribution shifts and performance decay.

Quantization reduces numerical precision (FP32 → INT8/INT4), decreasing memory usage and increasing inference speed with minimal quality loss.

Latency (token generation speed), GPU memory constraints, context window limits, and retrieval inefficiencies in RAG pipelines.

An agentic system can reason, plan, and execute actions using tools autonomously, rather than only generating text responses.

ReAct combines reasoning and acting by interleaving thought steps with tool usage, enabling multi-step problem solving.

Function calling enforces structured schemas and reduces ambiguity, making tool execution deterministic and production-safe.

They use shared state graphs or orchestration layers where agents communicate via structured memory and conditional execution flows.

Evaluation includes hallucination rate, factual consistency, instruction adherence, latency, and human preference scoring.

A stronger model is used to evaluate outputs of smaller models based on predefined rubrics like relevance, coherence, and factual correctness.

Prompt injection manipulates model behavior through malicious input in retrieved documents, potentially overriding system instructions.