<img height="1" width="1" style="display:none;" alt="" src="https://px.ads.linkedin.com/collect/?pid=6889170&amp;fmt=gif">
Skip to content
  • There are no suggestions because the search field is empty.

Nebula AI Case Explorer™ | AI 101: A Plain-Language Guide to AI and Nebula AI Case Explorer™

This article aims to create a working understanding of how AI Case Explorer's AI works, without a technical background. It covers the core AI concepts first, then maps each one directly onto a feature you use in the product.

The basics

Artificial intelligence (AI) is software that performs tasks — recognizing patterns, generating text, making predictions — that normally require human judgment. AI Case Explorer specifically uses generative AI: AI that produces new content (a summary, an answer, a score with reasoning) rather than just sorting or retrieving existing data.

Large language model (LLM) is the specific kind of AI behind generative AI tools like AI Case Explorer's chat and summaries. It's a model trained on enormous amounts of text to predict and produce language — which is why it can read a document and write a coherent summary, or answer a question in plain sentences, rather than just returning a list of keyword matches.

Model is the trained system itself — think of it as the "engine." AI Case Explorer runs on Microsoft Azure OpenAI's models, hosted on KLDiscovery's own secure infrastructure.

Training is the process of teaching a model by having it learn from massive amounts of text. This already happened before AI Case Explorer ever saw your documents — the underlying model was trained by Microsoft/OpenAI on general text, not on your case data. Your data is never used to train the model further — a distinction worth understanding, because it's the basis of the product's core privacy commitment.

Inference is what happens every time you actually use the AI — feeding it a question or a document and getting a response back. This is the only thing that happens with your data: inference, not training.

Talking to the AI

Prompt is the instruction or question you give an AI system. In AI Case Explorer, your case summary functions as a very detailed, structured prompt — it's the primary way you tell the AI what matters about your case. See Writing a Strong Case Summary.

Token is the basic unit an LLM reads and writes text in — roughly a word or part of a word (not the same as a "token" in security/auth contexts). Anywhere you see a size limit expressed in relation to how much text the AI can process at once, it's ultimately about tokens.

Context window is the maximum amount of text a model can consider at once — your prompt, the source material, and its own response all have to fit inside it. This is why very long documents or very long AI Chat conversations eventually hit a limit.

Hallucination is when an AI states something confidently that isn't true or supported by the source material — a well-known limitation of all generative AI, not something unique to this product. It's the single biggest reason AI Case Explorer is built the way it is. Every relevance score comes with a stated reason, every AI Chat answer provides a source document, and a persistent disclaimer reminds you to verify before relying on AI output. Treat AI output as a strong first draft or a well-informed pointer, not a verified fact, until you've checked the source.

How the AI works with your documents

Embeddings are a way of converting text into numbers that capture its meaning, so a computer can compare how similar two pieces of text are — even if they don't share any of the same words. This is what makes it possible to search "conceptually" rather than just by exact keyword match.

Vector search (sometimes called semantic search) is searching by meaning using embeddings, rather than by exact keyword. It's part of how AI Case Explorer's Basic AI Chat finds relevant passages in your documents even when your question doesn't use the same words the documents do. Agentic Chat will take this a step further and will leverage vector search if this is the best tool to use to answer your prompt.

Retrieval-augmented generation (RAG) is the overall technique of having an AI look up relevant material first, then generate an answer grounded in what it found — rather than generating an answer purely from what it learned during training. This is exactly what Basic AI Chat does: it retrieves passages from your document set, then writes an answer based on those passages, with citations pointing back to them.

Agentic AI describes an AI system that can plan and take multiple steps on its own to accomplish a goal — searching, checking its own results, refining its approach — rather than answering in a single pass. AI Case Explorer's Agentic Chat (currently in Beta) is exactly this: for a complex question, it can search, review what it found, and synthesize an answer across several steps, showing you its progress along the way. The standard Basic chat mode, by contrast, is closer to a single-pass RAG lookup — faster, and well-suited to direct factual questions.

Relevance scoring / classification is the AI reading a document and estimating how likely it is to matter to your case, based on the context you gave it (your case summary). AI Case Explorer's six-level scale (None, Highly Unlikely, Unlikely, Possibly, Likely, Highly Likely) is this in action — see Exploring Documents. Critically, the AI doesn't just output a score — it outputs its reasoning, so you can evaluate whether you agree.

Named entity recognition / information extraction is the general AI technique of pulling structured things — people, dates, events, organizations — out of unstructured text. This is what powers Key People and Key Events: the AI isn't guessing at a summary, it's specifically extracting and organizing named people and dated events from across your document set.

Keeping AI accountable

Human-in-the-loop or lawyer-in-the-loop describes a system designed so a person or lawyer reviews and can correct the AI's output, rather than the AI acting fully autonomously. AI Case Explorer is built around this: you agree or disagree with relevance scores, and that feedback measurably changes the product's next output — a genuine human-in-the-loop design, not just a UI gesture.

Feedback loop is what happens when the system you're correcting actually learns from the correction. AI Case Explorer's AI Relevancy feature is a concrete, visible feedback loop: your agree/disagree feedback (with reasons) gets turned into proposed changes to your case summary, which you approve, and which then improves the next analysis run.

Agreement rate is simply the percentage of time you agreed with the AI's relevance call, tracked as you review. It's useful as a citable, concrete measure of how well the AI is doing on your specific matter — not a vague assurance, but a number you can point to.

Version history means every prior state of your work is preserved rather than overwritten, so you can always see what changed and go back. See Project List and Version History.

AI Case Explorer Terms

A few of the terms that come up constantly in AI Case Explorer aren't AI concepts at all — they're eDiscovery or product terms that are easy to conflate with each other.

  • Custodian — the person or source a document was originally collected from (e.g., the mailbox owner an email was pulled from). This is an eDiscovery concept, not an AI-generated judgment about who matters — see the note on the Custodian tier in Key Themes, Key People, and Key Events. It's easy to conflate with Key Person (below), but they answer different questions: "where did this come from" vs. "who does the AI think matters."
  • Theme — a subject-matter grouping the AI identified (e.g., "Accounting Fraud") based on the case summary provided. Themes are about what a document is about. You can add, edit, or delete key themes during project creation or AI analysis revision.
  • Cluster — a grouping of documents based on topic similarity, visualized in Document Explorer's tree map and sunburst views. Clusters and themes can overlap but aren't the same thing — a cluster is a structural grouping that is undirected vs. a theme that is directed by your team via the case summary.
  • Storyline — a narrative thread of related key events over time, distinct from a theme (subject matter) or a cluster (document similarity).
  • "Key" anything (Key Person, Key Document, Key Event, Key Theme) — across the product, "Key" consistently means "the AI has identified this as likely significant to the case," as opposed to just present in the data. A custodian or a date on its own isn't "key" — being flagged as likely significant is what earns the label.
  • Relevance vs. responsiveness — in traditional eDiscovery, "responsive" usually means a document falls within a formal discovery request. AI Case Explorer's relevance score is a broader, AI-driven estimate of how much a document matters to your case theory as described in your case summary — related, but not a legal responsiveness determination.
Where AI Case Explorer draws firm lines
  • Your data is never used to train any model — for AI Case Explorer.
  • You own all inputs and outputs. The AI doesn't retain or share your data with other clients.
  • Nothing is a black box by design. Relevance reasoning, citations, agreement rates, and version diffs all exist for you to reference so you're never asked to just trust the AI's word.
  • AI output is a starting point, not a verdict. It's explicitly positioned as assistive — you're expected to apply professional judgment before relying on it.
Related articles