Tauseef Fayyaz

A to Z

Glossary

Every concept in this sheet, plus the smaller terms it uses. Select a concept to open its card.

Choosing chunk size: Small chunks match questions precisely but lose surrounding context; large chunks keep context but dilute the embedding and waste prompt space.
Chunk metadata: Fields stored with each chunk such as source URL, title, section heading, date, language, product version and access group.
Chunking: Splitting documents into passages small enough to embed and retrieve precisely, but large enough to make sense alone.

Document parsing: Extracting clean text from PDFs, HTML, slides and docs while keeping structure such as headings, tables and lists.

Fixed-size chunks with overlap: Cutting text every N tokens or characters, with each chunk repeating the last part of the previous one.

Indexing and query pipelines: Indexing runs ahead of time: load, clean, chunk, embed and store documents.

Knowledge base: The collection of documents the system can retrieve from: docs, tickets, PDFs, wiki pages or database rows.

RAG vs fine-tuning: RAG adds knowledge at question time; fine-tuning changes model weights to learn a style, format or narrow skill.
RAG vs long context: If all your content fits comfortably in the model context window, you can sometimes paste it in instead of retrieving.
Retrieval augmented generation: A pattern where your code first searches your own documents for passages relevant to a question, then gives those passages to an LLM to answer from.

Semantic chunking: Placing chunk boundaries where the meaning shifts, detected by comparing embeddings of neighbouring sentences.
Small-to-big retrieval: Embed small chunks for precise matching, but hand the model the larger parent section they came from.
Start with a naive baseline: Fixed-size chunks, one embedding model, top 5 by vector similarity and a simple grounded prompt.
Structure-aware chunking: Splitting on natural boundaries first, such as headings, then paragraphs, then sentences, and only cutting mid-sentence as a last resort.