Field notes / Glossary
Glossary - draft
Working vocabulary for how AI systems select, admit, and ground information—definitions only, shared across Field Notes.
This page lists definitions, not advice. For how to think about a problem, use a note; for what to build, use a pattern. Cross-links at the end wire this glossary back into the rest of the library.
This glossary is selective: it keeps terms that shape system behavior and interfaces—not an exhaustive catalog of AI or ML jargon.
These terms describe how information often flows through modern AI systems—especially systems that retrieve, transform, or generate content (including retrieval-heavy products, agents, and multimodal pipelines):
Retrieval and ranking use embeddings; they do not replace judgments about evidence, contracts, or failure behavior.
The total collection of source content your system can use.
A corpus is not the same as a clean index: ingestion gaps, duplicates, and drift still produce a bad experience at query time.
If the corpus is stale or messy, retrieval and generation quality drop no matter how strong the model is.
The process that selects which corpus fragments enter context for a task.
Think of retrieval as evidence selection, not just search ranking.
Retrieval is not relevance scoring alone—selection must answer what may legally and usefully enter the next step, not only what looks similar.
A bounded content unit used for indexing and retrieval.
Chunk boundaries should follow meaning and task intent.
Poor chunking splits concepts across units and degrades retrieval and answers even when the corpus and model are strong.
A chunk is not an arbitrary token window—if boundaries ignore semantics, evidence and grounding inherit that damage.
Linking model output to explicit source evidence.
Grounding improves traceability and lowers hallucination risk.
Grounding is not citation decoration—if outputs cannot be tied back to admissible evidence, you only have stylistic provenance, not a grounded system.
Information admitted into a reasoning or generation step after retrieval, filtering, or policy checks—not “everything the model once saw.”
Evidence quality is about fitness for the task and traceability, not raw token count.
Evidence is not the whole corpus—admission is a deliberate gate; “more text” is not a substitute for policy.
A vector representation of text (or other modalities) used for semantic retrieval and related ranking.
Embeddings compress meaning for search; they do not replace contracts, evaluation, or system behavior.
Embeddings are not understanding—similarity in vector space is a heuristic, not a guarantee of correctness or safety.
Compression: good systems select, admit, and ground evidence—not only retrieve and generate.