Understanding AI Responses

Archivist uses AI to search your documents and answer questions. This page explains how that process works in plain terms, so you can understand why the AI responds the way it does — and how to get better results.

How the AI Finds Answers

When you ask a question, the AI doesn't read through every word of every document from start to finish. Instead, it follows a faster process:

1. Your question is converted into a pattern that represents its meaning — not the exact words, but the concept behind them.
2. That pattern is compared against all the passages in your database (or just the ones in your selected file or file set).
3. The closest-matching passages are retrieved — the ones whose meaning is most similar to your question.
4. The AI reads those passages and writes a response based on what they contain.

This is why organizing your documents into file sets and choosing good chunk sizes matters — they directly affect which passages the AI finds and how much context it has to work with.

Why Answers Are Sometimes Incomplete

The AI can only work with the passages it retrieves. If the answer to your question is spread across a passage that wasn't retrieved, the AI won't know about it. A few common reasons this happens:

The question is too vague. A broad question like "Tell me about the project" might match many passages loosely instead of matching the right ones closely. Try being more specific: "What was the projected budget for Phase 2?"

The answer spans multiple sections. If the relevant information is in two different parts of a long document, and only one passage was retrieved, the AI might give a partial answer. You can increase the number of passages retrieved in AI Settings to cast a wider net.

The chunk size was too small or too large. If passages are very short, important context might be split across two passages. If they're very long, the matching might be less precise. See Uploading Documents for guidance on choosing a chunk size.

Why Answers Are Sometimes Wrong

The AI generates responses based on what it reads in your passages, but it's not perfect:

It can misinterpret context. If a passage contains a hypothetical scenario ("If the budget were doubled..."), the AI might present it as a fact. Always check citations when accuracy matters.

It can over-generalize. When the retrieved passages touch on related but different topics, the AI might blend them in a way that loses nuance.

It can't do math reliably. If you ask it to calculate totals from a spreadsheet or compare numbers across documents, double-check the results. AI language models are designed for language, not arithmetic.

Getting Better Results

Here are some practical ways to improve the answers you get:

Ask specific questions

Instead of "What does this document say?", try "What are the three main recommendations in the executive summary?" Specific questions match more precisely against specific passages.

Narrow your search

Select a single file or file set instead of searching everything. The fewer irrelevant passages the AI has to sort through, the better its results will be.

Adjust retrieval settings

In AI Settings, you can fine-tune how the AI searches:

• More passages — Increase "Max Chunks to Retrieve" if answers feel incomplete. The AI will consider more source material.
• Lower similarity threshold — If the AI is returning too few results, lower the threshold percentage so more passages qualify.
• Adaptive filtering — The recommended mode. It ensures the AI always has some passages to work with, even if none are a perfect match.

Re-chunk if needed

If you uploaded a document and the answers aren't great, consider deleting it from the Browse tab and re-uploading with a different chunk size. Larger chunks (800–1000 tokens) preserve more context per passage. Smaller chunks (256–512 tokens) give more precise matching.

Try rephrasing

If a question doesn't get good results, try asking it a different way. The AI matches meaning, not exact words, but different phrasings can emphasize different aspects of what you're looking for.

What the Settings Actually Do

Here's a plain-language summary of the key settings that affect answer quality:

Setting	What it does	When to change it
Max Chunks	How many passages the AI reads before answering	Increase if answers feel incomplete
Similarity Threshold	How closely a passage must match your question	Lower if too few results; raise if results feel off-topic
Filtering Mode	Whether to always return results or only when confident	Use Adaptive (default) unless you want strict quality control
Temperature	How creative vs. focused the AI's writing style is	Lower for factual lookups, raise for brainstorming
Chunk Size	How large each passage is (set during upload)	Larger for context-heavy docs, smaller for precise search
Show Citations	Whether to display source passages with each answer	Turn on when you need to verify accuracy