Image Tagging System — Intake Pipeline

Step 4 — How AI Tagging Works

Step 4 is where files get their meaning. By the time a file reaches this step it has already been converted, deduplicated, and hash-renamed — it's a clean, canonical file ready to be understood.

What happens

The vision AI receives each image and proposes two layers of tags: hr: product hierarchy tags (e.g. hr:pool-ramp, hr:stair-tread) and free-form content tags (e.g. dog, outdoor, lifestyle, in-use).
exiftool writes those tags directly into the file as XMP keywords — embedded in the file itself, not stored in a separate database.
The original filename and original path are also preserved in XMP fields before the hash rename, so nothing is lost.

The living knowledge base

Step 4 is not a one-shot operation. Every time the AI tags a batch, it consults two files that accumulate knowledge over time:

rules.md — distilled distinctions extracted from human corrections. For example: "nosing ≠ stair tread", "water + ramp + dog = pool context even if pool edge isn't in frame", "deck treads go on horizontal surfaces, not stairs". This file grows with every review cycle.
training-set.json — a library of verified image → tag pairs that the AI uses as worked examples. Seeded with 10 hand-tagged images on day one; every approved review batch adds more.

After each human review cycle, an agent reads the corrections, extracts new rules, and adds verified examples to both files. The next batch of tagging inherits that richer context automatically.

Why this matters: On day one, the AI is a general vision model with no knowledge of HandiRamp's product line. By cycle ten, it has a deep domain-specific memory — it knows your products, your photography style, and your edge cases. That compounding is what makes this system fundamentally different from a one-shot generic tagger. You are not just tagging files; you are building a proprietary classification engine that gets better the more you use it.

What it does not do

Step 4 does not determine whether a file is approved or final. It proposes tags. Human review (outside this pipeline) is what validates and corrects those proposals — and that feedback is what makes the next round of Step 4 smarter.