AI Coding

AI coding is the core feature of Paideias. It reads your documents, applies codes from your codebook, and returns passages with confidence scores and a short explanation. Your role is to review the suggestions — approving what fits and rejecting what doesn't. This page picks up right where Building a Codebook left off: the moment that page's codebook got its first definition, the two transcripts sitting queued since Uploading Documents started coding themselves.

Watch it happen: real coding, real review

Every screen below is real AI output against the actual Jordan and Marisol transcripts — not staged text. Click through it yourself, or let it play.

How AI coding works

When a document processes, Paideias runs it through a multi-step pipeline:

1. Chunking

Your document is broken into chunks — typically paragraphs or semantically meaningful segments, like the individual sentences from Jordan's answer that each got their own suggestion above. Each chunk is evaluated independently, so the AI focuses on one passage at a time without losing context.

2. Prompt construction

For each chunk, the AI receives the chunk text, your project's conceptual framework, your full codebook (every code name and definition), and instructions for how to evaluate and assign codes. This is why well-written definitions matter so much — the AI compares the passage against each definition to decide if there's a match, exactly as described in Why definitions matter.

3. Evaluation

The AI evaluates each code against the chunk and assigns a confidence score from 0 to 1 for anything relevant. In the walkthrough above, Jordan's passages scored between 0.88 and 0.95 against Ability — a genuinely strong, consistent match.

4. Return

The AI returns coded passages with the text, the suggested code, the confidence score, and a short explanation — like "Defines trust as evidence-based, directly linked to evaluating ability through actions," shown when approving above.

Running AI coding

Interpreting confidence scores

• 0.9 – 1.0: Strong match. Usually safe to approve — like the 95% suggestion approved above.
• 0.7 – 0.89: Moderate match. Relevant elements present, but review the passage carefully.
• 0.5 – 0.69: Weak match. The connection is tenuous; many of these get rejected.
• Below 0.5: Filtered out entirely — never shown, because the AI doesn't consider them relevant enough to surface.

What are emergent codes?

Sometimes the AI finds a passage that doesn't match any code in your codebook but clearly contains something important. It can flag an emergent code — a new code that doesn't exist yet. You can add it to your codebook, map it to something you already have, or dismiss it.

Reviewing and approving

The review step is where your judgment matters most. The AI identifies candidates; you decide what actually belongs in your analysis.

• Don't approve automatically. Even a 95% suggestion should be read against the real passage, the way we did above.
• Read the explanation. If the AI's stated reasoning doesn't hold up on a second look, reject it.
• Consider context. The AI evaluates chunks in isolation. You may know from the full document that a passage means something different in context.
• Approve in batches. On longer documents, review in groups to stay in an analytical flow rather than second-guessing every single click.

When a section couldn't be analysed

Occasionally an AI provider hiccups on part of a long document — a rate limit, a network blip. Paideias tracks this per section instead of failing the whole document, or worse, silently pretending everything was coded. You'll see a toast like "coded, but 2 sections couldn't be analyzed", and the document gets a small amber ⚠ N incomplete badge in the document list.

Running AI coding across multiple documents

You can run AI coding on individual documents (as above) or across your whole document set at once. Running across all documents keeps things consistent — every document is processed with the same codebook and prompts, which matters most before inter-coder agreement analysis, where every document needs uniform coding.

Re-running AI coding

The same button that says Run AI on a fresh document changes to Re-run AI the moment that document has any suggestions on it — same button, same click, the label just reflects what it's about to do.

Re-running is not destructive. It's a targeted refresh, not a wipe:

• Approved codes stay exactly where they are — both AI-approved and manually applied ones. Re-running never touches a decision you already made.
• Rejected and still-pending suggestions are cleared and replaced with a fresh AI pass against your current codebook.
• The document goes back into the queue with a processing status, exactly like a first run — see Why documents wait in queue if it seems to sit there.