Start from a Vague Request: Use Semantic Search to Find Deliverable Material

Semantic search is useful when the request is written in words rather than represented by a reference image.

1. Break Down the Request

Identify the subject, scene, color, style, and intended use. A vague request such as "warm product image" can become "wooden desk product photo, warm light, simple background".

2. Open Semantic Search

Enter the description in the semantic search dialog. Start with a clear but not overly long prompt.

Open Semantic in the workspace toolbar. The library needs semantic parsing first; this searches depicted content, while OCR/keyword search looks for text printed in images.

3. Set Scope and Result Count

Search broadly when you do not know where the material is. Narrow to a folder or collection when the request belongs to a known project. Increase result count when the prompt is exploratory.

Use Options to set the maximum count and choose the database, folders or collections. For inspiration, 30–100 candidates is a useful starting point; for client delivery, limit the scope to material approved for that project.

4. Review the First Round

Look for direction rather than perfection. Note which words worked and which results are irrelevant.

The results drawer and the left Results history keep the search available for review. Inspect subjects, scenes, colors and image quality rather than judging by filenames alone.

5. Narrow the Search Scope

If results are too broad, reopen Semantic → Options and choose the relevant folders or collection before searching again. The main file-list filter toolbar does not directly filter the open semantic-results drawer.

For additional text or quantity conditions, convert candidates into a collection, open that collection, and choose Actions → Filter Collection. This opens its own file/image, count, and text conditions. Ratings, color labels, and capture dates in the main file-list toolbar are a different set of controls.

6. Run a Second Prompt

Refine the wording based on the first result set. Add color, scene, object, or style details.

For a broad “product photo”, try “skincare bottles on a white background”. If “female model standing on the left against a blue background” is too restrictive, shorten it to “female model, blue background”. Add terms such as overhead, indoor, outdoor, flat lay or black-and-white when the style is wrong. Descriptions such as “sharp” or “no text” are search clues, not guaranteed quality filters; inspect the result.

7. Add Candidates to a Collection

Add promising results to a project collection so they can be reviewed and exported together. A collection does not move or copy the originals. The history entry's Actions menu adds the whole result, regardless of drawer selection. For a shortlist, use an explicitly selected-image action in the drawer. You can combine whole searches in one existing collection.

8. Export Delivery Material

History export acts on the whole result. For a subset, use an explicitly selected-image export in the drawer and verify the count.

Names export creates CSV with optional full paths; Images with Paths adds a TXT per image; spreadsheet output creates XLSX. Exported extracted images are not the whole source spreadsheet, PDF, or video. Confirm the source material and rights before delivery.

Common Cases

• Results are too broad: add subject and scene details.
• Results miss the style: add color or mood words.
• No results: check whether semantic data was generated and broaden the prompt.

Screenshot Example

The prompt 白雪覆盖的群山 means “snow-covered mountains”. In a sample library of four landscape photos and identical backups, the search returns eight images; the first two are the mountain original and backup at 40.63%. The input, options, and result screenshots belong to this search. The score is not a universal match threshold or a delivery-quality rating.