The AI reading glossary.
Plain-English definitions of the terms behind trustworthy document AI: what each one means, and why it matters for reading without hallucinations.
The most-connected terms in the glossary, and the natural entry points to everything else.
Abstractive summarization
Abstractive summarization rewrites the source in fresh words. It reads naturally but must be grounded in citations to stay faithful to the original.
AI hallucination
An AI hallucination is a false or unsupported claim delivered as confident fact. In document AI the fix is source-grounding: answering only from retrieved passages and citing them.
Chunking
Chunking splits a document into smaller passages that can be embedded and retrieved on their own. Chunk size shapes how well a system recalls the right text and how precisely it can cite.
Citation
A pointer from a statement to the source that supports it. In AI, a citation links a generated claim to the exact passage it came from - so the claim can be verified, not just trusted.
Context stuffing
Context stuffing dumps a whole document into the prompt. It's costly, hits context limits, and dilutes accuracy - retrieval feeds only the relevant passages instead.
Context window
A context window is the fixed amount of text a model can hold at once, measured in tokens. What doesn't fit is left out, which is why long documents are chunked and only relevant passages fed in.
Cosine similarity
Cosine similarity scores how closely two embeddings point in the same direction - the standard way semantic search ranks which passages are most relevant to a query.
Few-shot prompting
Few-shot prompting gives a model a handful of worked examples in the prompt to steer its format and behavior, without any retraining.
Fine-tuning
Fine-tuning further-trains a base model on extra examples to specialize it. For document tasks, retrieval-augmented generation is often a cheaper alternative.
Keyword search
Keyword search (lexical, BM25) matches the exact words in a query. It's precise for names and codes but misses synonyms - which semantic search handles.
Knowledge base
A curated store of documents a team relies on - wikis, runbooks, policies. Searching it is hard; pairing retrieval with citations turns it into something that actually answers.
Knowledge cutoff
A model's knowledge cutoff is the date its training data ends. Anything newer is unknown unless supplied at query time - the reason document grounding matters.
Prompt
A prompt is the instruction and context you give a language model. In document AI the prompt includes your question plus the retrieved source passages.
Prompt injection
Prompt injection hides malicious instructions inside content to hijack a model's behavior - a real risk whenever AI reads untrusted documents or web pages.
Reranking
Reranking is a second pass over search results: a slower, sharper model re-scores the top hits so the passage that best answers the query rises to the top.
Retrieval-augmented generation (RAG)
A technique where an AI retrieves relevant passages from a source first, then generates an answer grounded in them - reducing hallucination and enabling exact citations.
Scroll-to-source citation
A scroll-to-source citation is a live pointer: click it and the original document scrolls to and highlights the exact passage behind the claim, so verifying takes one click.
Semantic chunking
Semantic chunking splits a document by meaning - at topic boundaries - rather than fixed character counts, so each chunk is self-contained and retrieves cleanly.
Semantic search
Semantic search finds text by meaning, not exact keywords - using vector embeddings to match a query against passages that say the same thing in different words.
Source-grounding
Source-grounding constrains an AI answer to specific retrieved source text, with a citation on each claim, instead of letting the model answer from its own memory.
Temperature
Temperature controls how random a model's output is. Low values give focused, repeatable answers - what you want for faithful, citable document summaries.
Token
A token is the sub-word unit a language model reads and is billed by - roughly three-quarters of a word in English. Context limits and costs are measured in tokens.
Top-k retrieval
Top-k retrieval picks the k most relevant chunks of a document to feed the model, keeping prompts focused and grounded instead of stuffing in everything.
Vector database
A vector database stores and indexes embeddings so text can be found by meaning, not exact words. It powers fast semantic search over large document sets.
Vector embedding
A vector embedding is a list of numbers that captures what text means, so passages with similar meanings sit close together in space. Embeddings power semantic search and retrieval.
Stop digging. Start asking.
Add Sidenote to your browser, open any page in your wiki, and ask it the question you’ve been Slacking the team about.
7-day Pro trial · No card required · Free plan forever