Does this search what people say in the ads?

No. This template indexes what is on screen, who is in frame, and the ad copy you supply as a field. Spoken audio is not part of it, and the measured corpus carries no transcripts at all, so nothing on this page depends on speech.

What do the relevance scores look like?

Cross-modal text-to-video similarity sits in a narrow band, roughly 0.11 to 0.13 on the measured corpus. That is normal for this kind of retrieval and it is why the template does not render a score as a percentage: an 11% label reads as a broken product when the ranking is working correctly. Use rank order, or a bar relative to the top result.

Why are faces a separate collection instead of a field on the scene?

One scene can contain several faces and one face appears in many scenes, so attaching faces to scenes would force a choice about which face the scene is. Keeping them separate lets a face query return every appearance across the archive and a scene query stay about composition.

Do the clusters work out of the box?

Clustering is a per-brand run rather than a step in the pipeline, and on the measured namespace 9 of 16 runs completed, 3 failed outright, 1 was in progress and 1 completed with errors. Plan for a brand's cluster run needing a retry, and read the cluster status before presenting themes as complete.

How much of the archive does copy-search reach?

Ad copy is a field you supply on the object, not something the pipeline extracts from the video, so coverage is whatever your export carries. On the measured corpus it is 38 creatives out of 271, about 14%, with a median chunk of 184 characters. The retriever is correct and the index is the real multilingual one, and it is worth knowing that copy-search answers over the creatives that have copy rather than over the whole archive.

Can I point it at a public ad library instead of my own exports?

The pipeline does not care where the video came from, only that the object carries a video URL and the copy. What changes is the shape of the media: the measured corpus is vertical ad creative, 5 to 80 seconds, which cuts into scenes readily. Longer single-take footage will produce one scene per file and the numbers on this page will not reproduce.

Advertising

Template v1.0 · updated 2026-09-11

Creative DNA for Ad Archives

Search every scene of every ad you have run, and find the ones that keep earning spend

An ad archive you can query by what is on screen. Every creative is cut into scenes at ingest, every face is indexed, and the ad copy is searchable beside them, so a question like 'the hook where someone holds the product to camera' returns the moment rather than the file.

Deploy starter pipeline

Performance marketing and creative teams sitting on an ad archive nobody can search

Studio will also guide you to Your own ad exports for zero-setup evaluation, then to a continuous data connection for production.

2,306

scenes across 271 source videos

Measured 2026-09-11 by paging the live ads-scenes collection (col_a96402bd17) to exhaustion. Median 5 scenes per video, mean 8.5, max 19; 39 of 271 videos yield exactly one scene.

8,168

face documents from the same archive

Live count on col_6cdf041caf, face_identity_extractor@v1 with scrfd_2.5g detection and quality scoring on.

brands in the corpus

13 distinct brand_slug values, of which four are re-mints of the same brand under a -2 suffix, so nine brands by name.

creatives carrying ad copy, out of 271

Live count on ads-text (col_e682a390cd), measured 2026-09-11: 38 documents, 38 distinct ad_id, one chunk each, median 184 characters. Copy is supplied per creative rather than extracted, so coverage is whatever your own export carries. On the measured corpus that is 14%.

2.5s

median scene length

At scene_detection_threshold 0.5 on source clips whose own median length is 29 seconds. Mean scene length 3.6s, max 29.5s.

What it looks like

Frame in, decisions fire, a verdict lands, a reviewer's call feeds back.

What goes in

ad-creatives
One object per creative: the video plus the ad metadata. ad_id, brand_slug and brand_name are required; the copy, headline, source and publish date are optional and travel with it.

Where the decisions fire

Illustrative frames; boxes show the decision path, not live model output

Two people filming a piece to camera with a ring light on a table — ads-scenes and ads-faces, both fed from the same object · A talking-head hook. The scene embedding carries the composition and setting; the face extractor indexes the person separately, which is what makes 'every ad this creator appears in' a single query.Still: Pexels / ANTONI SHKRABA production

A hand holding a dropper bottle close to the camera with a person behind it — ads-scenes · The product-to-camera beat that performance creative returns to. Scene-level indexing is what lets you ask for this shot across brands instead of scrubbing the ads you happen to remember.Still: Pexels / SHVETS production

Two people running along a tree-lined path at sunrise, seen in silhouette — ads-scenes, then clustering · A lifestyle establishing shot. Scenes like this are what cluster into a brand's recurring visual themes, which is the view an archive cannot give you file by file.Still: Pexels / David Kanigan

What comes out

ads-scenes
Scene-level documents with a multimodal embedding, timing, and the brand and ad they came from.
ads-faces
Face identity vectors, one document per detected face, with a quality score.
ads-text
Chunked ad copy, searchable on its own or beside the visual match. Populated only for creatives whose export carries copy.

How the namespace is wired

1 buckets, 3 collections, 0 clean views, 3 retrievers. The diagram generates the manifest below; they cannot drift apart.

SourceBucketCollectionClean viewRetrieverClick a node to inspect it

Reward signals

How reviewer decisions move the thresholds

Thresholds at ingest drift as the corpus changes. The reviewers working the queue are the ones who see where a threshold is wrong first, so this template routes their decisions back into the model that set it.

DocsInteractions API

Nothing in this template writes interaction signals back, and it is worth saying so rather than implying a flywheel that is not wired.

Explicit signals

None. No thumbs, no saves, no relevance marks are recorded by the template as shipped.

Implicit signals

None from the application. Mixpeek records retriever executions server-side in mxp_retriever_executions, which is telemetry about queries rather than feedback about results.

Where they land

mxp_retriever_executions

System collections in your namespace, on the same vector store as the rest of the template. They are yours to query.

How the loop closes

Adding a learning loop means posting to /v1/retrievers/interactions when a user opens a result. That is a deliberate next step, not something this template does for you.

DocsManifest schema

One file spins up the namespace. Generated from the diagram above. Also served at /templates/creative-dna.namespace.yaml.

# creative-dna: one manifest spins up the namespace.
# Platform manifest schema (GET /v1/discovery/schema). Validate with POST /v1/manifest/validate,
# apply with POST /v1/manifest/apply or the Deploy button. Wiring comes from the flow diagram:
# edges are bucket -> collection sources, collection -> retriever scope, retriever -> view.
version: '1.0'
metadata:
  name: creative-dna
  description: "Namespace template creative-dna. Generated from the flow diagram on mixpeek.com/templates/creative-dna."
namespaces:
  - name: creative-dna
    description: "Everything below lives in this namespace."
    feature_extractors:
      - name: multimodal_extractor
        version: v1
      - name: face_identity_extractor
        version: v1
      - name: text_extractor
        version: v1

# Data sources. A storage connection carries credentials, so it is created in Studio
# (or POST /v1/organizations/storage-connections) and synced into the bucket named here.
#   ad-library-pull: manual, one-shot, your ad creatives and their copy -> bucket ad-creatives
buckets:
  - name: ad-creatives
    namespace: creative-dna
    description: "Fed by ad-library-pull (manual, one-shot)."
    schema:
      properties:
        ad_id:
          type: string
        brand_slug:
          type: string
        brand_name:
          type: string
        source:
          type: string
        source_url:
          type: string
        headline:
          type: string
        primary_text:
          type: string
        ad_type:
          type: string
        creative_asset_url:
          type: video
        published_at:
          type: string
collections:
  - name: ads-scenes
    namespace: creative-dna
    description: "multimodal_extractor@v1 over bucket ad-creatives. Feeds brand-scenes."
    source:
      type: bucket
      bucket: ad-creatives
    feature_extractor:
      name: multimodal_extractor
      version: v1
      input_mappings:
        video: creative_asset_url
      field_passthrough:
        - source_path: ad_id
          required: true
        - source_path: brand_slug
          required: true
        - source_path: brand_name
          required: true
        - source_path: ad_type
        - source_path: source
        - source_path: source_url
        - source_path: headline
        - source_path: primary_text
        - source_path: published_at
    enabled: true
  - name: ads-faces
    namespace: creative-dna
    description: "face_identity_extractor@v1 over bucket ad-creatives. Feeds cast-identification."
    source:
      type: bucket
      bucket: ad-creatives
    feature_extractor:
      name: face_identity_extractor
      version: v1
      input_mappings:
        video: creative_asset_url
      field_passthrough:
        - source_path: ad_id
          required: true
        - source_path: brand_slug
          required: true
        - source_path: brand_name
          required: true
        - source_path: ad_type
        - source_path: source
        - source_path: source_url
        - source_path: headline
        - source_path: primary_text
        - source_path: published_at
    enabled: true
  - name: ads-text
    namespace: creative-dna
    description: "text_extractor@v1 over bucket ad-creatives. Feeds copy-search."
    source:
      type: bucket
      bucket: ad-creatives
    feature_extractor:
      name: text_extractor
      version: v1
      input_mappings:
        text: primary_text
      field_passthrough:
        - source_path: ad_id
          required: true
        - source_path: brand_slug
          required: true
        - source_path: brand_name
          required: true
        - source_path: ad_type
        - source_path: source
        - source_path: source_url
        - source_path: headline
        - source_path: primary_text
        - source_path: published_at
    enabled: true
retrievers:
  - name: brand-scenes
    namespace: creative-dna
    description: "Searches ads-scenes across 1 feature index."
    collections:
      - ads-scenes
    input_schema:
      query:
        type: text
        required: true
        description: "What to look for; searched across every index below"
    stages:
      - stage_name: search
        stage_id: feature_search
        parameters:
          searches:
            - feature_uri: "mixpeek://multimodal_extractor@v1/vertex_multimodal_embedding"
              query:
                input_mode: text
                value: "{{INPUT.query}}"
              top_k: 100
          fusion: rrf
          final_top_k: 100
    tags:
      - template:creative-dna
  - name: cast-identification
    namespace: creative-dna
    description: "Searches ads-faces across 1 feature index."
    collections:
      - ads-faces
    input_schema:
      query:
        type: text
        required: true
        description: "What to look for; searched across every index below"
    stages:
      - stage_name: search
        stage_id: feature_search
        parameters:
          searches:
            - feature_uri: "mixpeek://face_identity_extractor@v1/insightface__arcface"
              query:
                input_mode: text
                value: "{{INPUT.query}}"
              top_k: 50
          fusion: rrf
          final_top_k: 50
    tags:
      - template:creative-dna
  - name: copy-search
    namespace: creative-dna
    description: "Searches ads-text across 1 feature index."
    collections:
      - ads-text
    input_schema:
      query:
        type: text
        required: true
        description: "What to look for; searched across every index below"
    stages:
      - stage_name: search
        stage_id: feature_search
        parameters:
          searches:
            - feature_uri: "mixpeek://text_extractor@v1/multilingual_e5_large_instruct_v1"
              query:
                input_mode: text
                value: "{{INPUT.query}}"
              top_k: 50
          fusion: rrf
          final_top_k: 50
    tags:
      - template:creative-dna