In words
What it is, why it matters, and what it is like.
Why am I learning this?
This concept is your first step into retrieval-augmented generation (RAG) — the machinery that lets large language models answer questions using documents they have never memorised. Once you master admissible evidence spans, you can: build citation systems that make every answer checkable (the 'Citations and Evidence' topic in your library), design retrieval pipelines where provenance is tracked from ingest to answer, and later move on to advanced RAG work like agentic retrieval and GraphRAG. In any regulated setting — finance, law, medicine — an LLM that cannot show its sources is useless; this concept is the foundation for making an AI answer defensible.
The idea, in plain terms
Imagine you are a junior analyst at a financial firm. A senior partner asks you a question: 'What were the risks mentioned in the last quarterly filing of Company X?' You open the firm's document system and search. You pull up several documents, each with a few paragraphs that might answer. But you cannot just paste the first thing you find into your reply. You must point to the exact paragraph — the precise span of text — that supports each claim you make. That exact paragraph is an 'admissible evidence span'. It is admissible because it comes from a verified source (the filing), and it is a span because it is a specific, bounded section of text, not the whole document or a vague reference.
Now think about how a modern AI assistant works. When you ask a question, the AI does not just 'remember' the answer. It (or the system around it) retrieves relevant documents, finds passages, and generates an answer. If the answer is grounded in admissible evidence spans — meaning each claim is backed by a specific, traceable passage — then you can click a number in the answer and see the exact sentence that supports it. If the passage is not admissible — say it comes from an unverified blog or a stale version of the filing — the whole answer becomes suspect. In regulated settings, the difference between 'admissible' and 'inadmissible' can be the difference between a compliant answer and a lawsuit.
An analogy
Think of a courtroom trial. A lawyer cannot just say 'the defendant was at the scene' — she must present evidence: a witness who saw it, a CCTV clip, a signed document. That evidence must be admissible: it must be authentic (not fabricated), relevant (it actually speaks to the case), and properly obtained (not hearsay or illegally gathered). And it must be specific — you cannot cite 'the whole investigation file'; you point to exhibit A, page 3, lines 12–14.
An AI system that answers questions is like a lawyer building a case. The 'admissible evidence spans' are the specific passages it may cite. The system's retrieval step is like the lawyer's investigator pulling relevant documents. But the lawyer must check admissibility: is this passage from a trusted source? Is it about the question? Is it specific enough to pinpoint? Only if it passes these checks can it be used in the final argument — the AI's answer.
Where the analogy stops working: a courtroom has a judge who rules on admissibility. In AI, the admissibility check is often automated — a piece of code decides whether a passage is eligible based on metadata (source, date, access level) and relevance scores. Also, in court, evidence is physical or testimonial; in AI, evidence is text (or numbers, images) stored in a database. The crucial difference: the decision to admit evidence in court is made by a human expert with judgement; in AI, it is a rule-based filter, which means we must design those rules carefully.
Definition
An admissible evidence span is a specific, contiguous passage from a verified source document that an AI system is permitted to cite as support for a claim in its answer, satisfying criteria of source authenticity, relevance, and specificity.
Where this sits
You have not yet mastered any prerequisites formally, but this concept connects to several topics in your library that you will learn soon. It builds on the idea of retrieval-augmented generation (your parent concept), and it is closely tied to 'Citations and Evidence' — ensuring that claims point to supporting passages. It also relates to 'Chunking' because how you split documents into chunks determines what can later be an admissible span. And it is a prerequisite for 'Context Precision and Recall', which measure whether the retrieved context includes the admissible evidence. In your notes from *Large Language Models in Finance*, the key claim is that in finance, answers must be traceable to admissible evidence spans in source filings — making citation a compliance requirement, not a nicety. That is the heart of this concept.
