Library Indexing

Library Indexing is Nephele's AI analysis feature for your asset library. It scans your assets one by one, automatically applies content tags and style tags, and builds a cross-language semantic search index. Every model runs locally, and your images are never uploaded to the cloud.

What the index contains

Library Indexing isn't a single "tagging" step — it's three independent local AI modules working together:

1. Content recognition

By default it uses the PixAI Tagger v0.9 model (EVA02 architecture, running locally via ONNX, about 1.2 GB). For each image it outputs two kinds of tags:

The default confidence threshold for content tags is 0.35, and 0.85 for character tags. The higher the threshold, the more conservative but more accurate the tags will be (PixAI's bundled thresholds.csv overrides these two defaults).

The model automatically handles images with transparent backgrounds (compositing them onto a white background) and pads each image into a square before feeding it into the 448x448 input layer. Accelerator priority: DirectML (Windows) > CUDA > CPU.

In the same tagging pass, PixAI also produces a 1024-dimensional image embedding vector (in EVA02 space), stored alongside the tags for free, for use by reverse image search to find visually similar images.

2. CLIP style classification

Uses the CLIP ViT-B/32 model for zero-shot style recognition, matching against 18 predefined styles:

watercolor, oil painting, cel shading, flat color, pixel art, sketch, lineart, ink wash, 3D render, photograph, comic panel, realistic digital, pastel, impasto, vector art, concept art, chibi, monochrome

Each image takes the top 2 style tags by confidence (with a minimum threshold of 0.20). CLIP also powers text-to-image search (in a 512-dimensional joint text-image space). Style tags and content tags are two independent systems.

3. Multilingual semantic search

Uses gte-multilingual-base to encode both the tags and your query into the same semantic space (768 dimensions). This way, searching for "blue-toned girl" in your own language can still match images in your library that only have English Danbooru tags — it searches the semantic meaning of tags, then filters down to the tags that actually exist in your library. This model is an optional download, fetched on demand the first time you run a semantic search.

Tags become localized

During tagging, the English tags the model produces are replaced with their corresponding localized display names and written directly back to the Eagle element (only tags without a localized name keep their English form). In other words, to convert an existing library, just run the index again (force re-tag). Nephele records the source of every tag under the hood, so a re-tag only replaces the tags it wrote itself last time — it never touches the tags you added manually.

How to start indexing

Option 1: Through the interface

Open the "Library Indexing" view (from the left navigation bar or the Astrolabe wheel) and you'll see a masonry grid displaying the assets in your Eagle library. At the top there's a filter and action bar:

• Folder filter — process only the assets under a specific folder
• Extension filter — png, jpg, psd, svg, gif, webp
• Status filter — All / Indexed / Not indexed
• Processing limit — unlimited by default, set your own

After you click "Start indexing," Nephele first counts the unindexed assets that match your criteria, then processes them one by one.

Option 2: Let the Agent do it

Just tell Nephele:

• "Batch-tag my Eagle asset library"
• "Add AI tags to the images in the 'Character Reference' folder"
• "Process all the illustrations that haven't been indexed yet"

The Agent will automatically call the indexing tool and report the results when it's done.

Processing parameters

The confidence threshold is the parameter most worth tuning. If you find the applied tags too noisy, raise it to 0.45 or 0.50; if you feel it's missing some obvious elements, lower it to 0.25.

Resuming after interruption

When a batch run starts, Nephele writes a checkpoint file locally (~/.nephele_workshop/tagger_state.json) recording the current task scope. If the software crashes or is force-quit mid-process, the next time it starts Nephele detects this uncleared checkpoint and asks whether you want to continue your previous task.

After a normal completion or a manual cancellation, the checkpoint is cleared automatically.

Deduplication and skipping

Nephele won't reprocess assets that have already been indexed. The criterion is whether an asset carries the Nephele-tagged marker tag:

• Before a batch run, already-marked assets are pre-filtered out automatically
• In direct mode, by default only "untagged" assets are processed (unless you've specified other tag filter conditions)
• If you deliberately want to reprocess a particular image, you can manually delete its Nephele-tagged tag

How it reads and writes

Nephele reads from and writes to Eagle's .library/ filesystem and metadata.json directly, without going through the Eagle app's HTTP API. As long as you've configured the Eagle library path, it can process your assets — Eagle itself doesn't need to be running in the background. During processing it automatically filters out non-image files (video, audio, archives, etc.), handling only png, jpg, jpeg, gif, webp, bmp, tiff, tif.

Single-image preview

Don't want to run a batch — just want to see what tags a particular image would get? In the Library Indexing view, click any asset and the right-hand panel shows a "Preview tags" button. After you click it, Nephele analyzes that image and displays the predictions, but does not write anything to Eagle.

The preview results include: content tags, character tags, and style tags.

After indexing: searching your library

Once indexing is complete, an asset's tags and style are written back into Eagle's metadata (metadata.json), while the image embedding vectors are stored in a local cache (image similarity uses the EVA02 tagger space, text-to-image search uses the CLIP space; the two sets of vectors are stored separately and never mixed). You can put this data to use in the following ways:

Agent conversation

Just tell Nephele:

• "Search for images tagged 'weapon'"
• "Find all illustrations rated 4 stars or higher"
• "Show me which PNGs are in the 'Character Reference' folder"
• "Give me the full information for this image"

The Agent automatically combines multiple parameters for precise filtering. For example, if you say "find 4-star-and-up JPGs tagged mecha," it passes star=4, tags=["mecha"], and ext="jpg" all at once.

Manual search in the toolbox

The toolbox has a "Library Search" entry where you can manually set keywords, tags, star ratings, and other conditions before running the search.

Search parameters

Search results

Each returned asset includes: id, name, ext, tags, star, width/height, filePath, thumbnail, folders.

Other index-driven features

• Similar image search — give it a reference image and find the visually closest assets in your own library (based on the EVA02 tagger embedding vector, with tag Jaccard as a fallback)
• Text-to-image search — describe the image you want in a sentence and retrieve the most semantically matching images in your indexed library (based on CLIP embeddings, no exact tags needed). CLIP is trained on English, so a non-English description is first translated into English by the Agent before searching
• Agent auto-organizing — "Move the images tagged 'mecha' into the 'Mecha Reference' folder"

If you haven't built an index yet, these features will be substantially limited. We recommend running a full index on your most-used library first.