This article is intended for users who are unfamiliar with using AI tools. It details the nature of chats and the building blocks of good prompt engineering.
What LOIS is Not
It’s easy to assume LOIS works like something familiar when it doesn't. Before going further, let’s break down how LOIS and other AI tools are different from search bars, like Google’s or Filevine’s search.
| Search bar | LOIS | |
|---|---|---|
| What it does | Finds records that contain your keywords | Calculates answers from structured data |
| Tolerance for vagueness | High — "close enough" usually works | Medium — it interprets your intent, but precision matters |
| What you get back | A list of matching results | Depends on how you asked — could be a number, a list, or a grouped table |
| When it fails | Returns nothing or unrelated results | May return something that looks right but isn't — it won't always tell you it guessed |
Note
That last row is the most important thing to internalize: unlike a search bar that returns zero results, LOIS will return something — it just might not be exactly what you meant. This is why how you phrase your question matters more than you might expect.
What LOIS Is
It may help to use the following analogy: imagine there's a massive filing room where every piece of your firm's data lives—but it's all labeled in a language only database engineers speak. You don't speak that language, and that's fine. LOIS is your interpreter. You describe what you need in plain English, the interpreter translates it into the filing system's language, retrieves the records, and hands them back to you. But if you say "pull the cases that aren't going well," the interpreter has to make a judgment call about what "not going well" means in the filing system's terms. They might ask. They might guess. The more precise you are, the less they have to guess—and the more accurate what comes back will be.
Starter AI Terminology
| chat | A chat is a back-and-forth conversation between a user and an AI assistant. Each chat constitutes a session in which user messages and assistant responses build on one another, allowing the assistant to maintain continuity across turns within that session. |
| input or prompt | The input or prompt is a message a user sends to the AI. The message may take the form of a question, an instruction, a document submitted for review, or any combination of these. |
| output or response | The output is the reply produced by the AI after processing the prompt. Responses may take many forms, including plain text, drafted documents, summaries, tables, citations, or the results of tool calls (such as data retrieved from a case). |
| context window | The context window is the amount of information the assistant can consider at one time when generating a response. It includes the current prompt, recent conversation history, retrieved documents or data, and the assistant's own instructions. When a conversation or document grows very large, older or less relevant material may fall outside the window and no longer influence subsequent responses. |
| knowledge base | A knowledge base is the pool of information the AI can draw on to answer questions. In Filevine, these include project documents (such as medical records, pleadings, and correspondence), structured project data (contacts, deadlines, notes, billing, and calendar events), and case-law sources available on request. Information is retrieved from these sources on demand rather than stored in advance, and citations back to the source are provided so that any fact can be verified. |
Chatting
AI agents operate based on chats. A chat is a conversation thread between you and the AI agent, as the AI agent may request additional information, and you may iterate on the response it provides.
Once you begin chatting, you can find your chat history in the sidebar panel on the left.
When to Start a New Chat
You should start a new chat, rather than continuing the old chat, whenever you are switching to a new task, or if you have reached a dead end in the current chat.
Why start a new chat?
- Avoid context confusion. The AI uses the chat thread as context for its response. Making a new, unrelated request in the same chat can cause confusion when the AI attempts to use the rest of the chat thread as context.
- Keep speed and focus. More context in the conversation means more information for the AI to take into account. Long conversation can slow down responses.
- Start fresh. If you aren’t getting the response you want, or you want to compare responses for two different prompts, you can try starting a new chat without the prompt that isn’t working.
Managing Multiple Chats
Once you enter a prompt. the AI will begin “thinking” of its response. While the AI thinks, you’ll be unable to add or adjust the prompt, so as now to interrupt the AI. However, you can freely navigate away from the chat while the AI thinks, or even start another, separate chat.
Prompting
Think of prompting as a skill that can be developed with practice. In general, with prompting, be specific with your context, and don’t be afraid to iterate and explore.
Be Precise and Provide Context
The more precisely you can explain what you want, the better your result will be. Narrowing by date range, project type, or status reduces query time and keeps results focused. Instead of "Show me cases," try "List all Personal Injury cases in Active status created after January 1, 2025."
The more specific your question, the more accurate the result. Vague questions produce vague answers—just like they would with a real analyst.
Specify Expected Output
Like providing a date range, specifying the type of output you need can help you get the response you want. Are you looking for a count? A list? A breakdown by category? If you’re looking for a table, name the columns you want. If you need specific fields in the output, ask for them explicitly: "Include case name, assigned attorney, and open date."
Look Out for Aggregate Wording
A common source of confusion is aggregation. When you use words like "how many," "total," or "average," LOIS runs a calculation that collapses all your data down to a single summary value. You won't get the individual records—just the answer to the math question.
For example, if you ask "How many PI cases were opened this year?" or "What is the total billed across all active cases?," LOIS responds with the aggregate count: “47” or “$284,500.” If you want the list instead, try asking LOIS to "List all PI cases opened this year" or "Show me billing totals for each active case.”
Neither form is wrong—they answer different questions. The key is knowing which form you need before you ask.
Be Positive
Use positive framing, telling the AI what to do, rather than what to avoid.
Iterate and Follow Up
Getting to your desired output may involve some back and forth between you and LOIS—that's normal. A human analyst can't read your mind, and neither can LOIS. If your question is ambiguous, references data that isn't in the model, or could mean multiple things, LOIS may ask a clarifying question, return unexpected results, or explain why it couldn't find what you were looking for.
You can also provide feedback to LOIS to fine-tune the response—ask it to revise the response with a different output type or tone, or give more specific instruction to narrow down the scope.
Tip
Ask LOIS itself for help refining results. If the results look off, ask: "Why did you look at that table?" or "How did you interpret my question?" LOIS can walk you through its reasoning, just like an analyst explaining their work.
Seeing It in Action: Prompt to SQL
You don't need to read SQL to build skills at prompt engineering. Just notice how the words you choose shape the response LOIS provides.
| Your prompt | What the AI runs | Why it works this way |
|---|---|---|
| "Show me active cases" | SELECT all columns FROM projects WHERE status = 'Active' | No column filter specified — you may get more fields than expected. Name the columns you want. |
| "How many active PI cases are there?" | SELECT COUNT(*) FROM projects WHERE status = 'Active' AND project_type = 'Personal Injury' | COUNT returns one number. You won't see case names, attorneys, or any other detail. |
| "Cases that aren't going well" | Agent must guess what column or value means "not going well" | No field in the database is labeled this way. The agent will make a judgment call — results may vary. |
| "PI cases in Discovery opened after Jan 1, 2025 with no billing entries — show case name, attorney, and open date" | Clean, precise SQL with filters, an exclusion join, and specific columns selected | Every part of the request maps directly to a field or condition. This is the ideal form of a prompt. |
Comments
0 comments
Article is closed for comments.