Damsel

Multi-modal metadata generation for a feature film.

Damsel, decoded: building a search taxonomy from the story itself

Netflix wanted to explore the best ways to allow for enhanced recommendation and search leveraging language models. Members should be able to search not just for titles but moods, storylines, memorable scenes; saying “A dragon movie that actually scares me” should surface accurate results. The traditional metadata, contextual detail for a piece of content, was written mostly as basic tags; genre, cast, a synopsis, type of music. This desciptors don't provide enough contenxt for language models to be useful in improving search and recommendation.

Search can only find what the metadata knows is there.

Netflix came with two questions. How could search connect what members describe to what a title actually contains? And how could the feelings each scene elicits be matched to member cohorts to surface better recommendations? We ran Damsel through Arc View, Pilotly's content intelligence system, to show what a title looks like when its metadata comes from the film itself. Built scene by scene and moment by moment, the building blocks of the story become the metadata set a language model can interpret.

what are all the storylines?who carries it?which moment sells it?

One film, broken into scenes, and then into story beats

37

scenes

193

moments

110 minutes, every frame7 named characters, Elodie on screen for 30% of runtime6 storylines traced scene to scene

Arc View watched all 110 minutes of Damsel and broke it into 37 scenes and 193 moments. Each moment carries a narrative description, a story beat that names its characters, a video technical description of how it was shot, an audio technical description of how it sounds, and emotion tags. Above that sits a character layer, built from face detection and named against the cast, and a story layer that describes each scene with its characters by name and rolls scenes up into storylines.

From those layers we proposed a search taxonomy in four tiers: title, storyline, scene and moment. Each tier answers a different kind of query, and every description points back to a timecode.

Every description is tied to a timecode range.

The taxonomy, four tiers

TitleGenre, emotions, cast, themes
Storyline6 storylines traced across the film
Scene37 scenes, each with a story description that names its characters
Moment193 moments, each with a narrative, video technical and audio technical description, emotion tags and a story beat

One moment, a dozen ways to find it.

This is one of Damsel’s 193 moments. Tap a layer below the image to see what Arc View knows about it.

Scene 20 of 37Cave: Dragon EncounterMoment 4 of 836:26 to 36:36

Elodie in a ragged dress, looking up into the dark beside a bare, twisted branch — Before the reveal

A glowing amber eye with a slit pupil, open in the dark rock — The reveal

Narrative

A slow look across the cave wall ends when a glowing eye opens in the dark, revealing a hidden, ancient presence.

The scene unfolds in a dark, rocky cave, where the tension is palpable. The camera slowly pans across the rugged terrain, highlighting the intricate details of the rocks and the subtle shadows cast by the flickering light. Suddenly, a pair of glowing eyes emerges from the darkness, piercing through the gloom with an otherworldly intensity. This moment marks a pivotal shift in the narrative, as the audience is drawn into the unknown, hinting at a hidden danger or an ancient presence lurking in the shadows. The visual composition emphasizes the contrast between the dimly lit cave and the radiant eyes, creating a dramatic and suspenseful atmosphere that sets the stage for the unfolding events.

Story beat

First sight of the dragon. Elodie, alone in the cave: a gasp echoes off the rock, and a glowing yellow eye opens in a crevice of the stone.

The scene opens with a claustrophobic close-up on a dark, rugged rock face, immediately establishing an ominous atmosphere as a sharp, heavily reverberated gasp echoes through the cavernous space. Under dim, flickering light, the stationary camera captures the rough textures and crevices of the stone, where small patches of moss cling to the cold, shadowed surface. Accompanying this visual stillness is an ultra-low-energy, ambient C minor drone that rumbles quietly alongside the slow dripping of water, building a tense, imperceptible pulse at one hundred and eighteen beats per minute. As a subtle shift in the lighting occurs, the slow visual exploration of the stone texture takes a dramatic turn when a glowing yellow eye suddenly emerges from within a deep crevice of the rock. The camera shifts its focus entirely onto this radiant, otherworldly eye, which pierces through the gloom and serves as the stark, central point of contrast against the surrounding darkness. Synchronized with this unsettling visual reveal, the spatial audio tightens around the listener as an ominous swell rises, heightening the protagonist's immediate sense of vulnerability. This striking juxtaposition of the bright, ancient eye against the shadowy, suffocating abyss solidifies the presence of a hidden danger lurking just out of sight, leaving the audience suspended in a state of absolute dread.

Video technical

How it was shot: a stationary close-up on rough, dark stone under dim light. The light shifts to reveal a glowing yellow eye, and the focus moves onto it, the central point of contrast against the dark.

The video begins with a close-up shot of a rocky surface, characterized by its rough texture and dark, muted colors. The lighting is dim, casting shadows that accentuate the contours and crevices of the rocks. The camera remains stationary, focusing on the details of the rock's surface, which includes small patches of lighter-colored material, possibly lichen or moss. As the scene develops, there is a subtle shift in lighting, revealing a glowing yellow eye within the rock. This eye appears to be part of a larger, unseen creature, adding an element of mystery and intrigue. The focus then shifts to this glowing eye, which becomes the central point of interest. The lighting highlights the intricate details around the eye, creating a stark contrast against the dark background. The video wraps up with a continued emphasis on the glowing eye, maintaining the same dim lighting and shadowy atmosphere, leaving viewers with a sense of curiosity about the creature behind the eye.

Audio technical

How it sounds: an ultra-low-energy ambient drone in C minor at 118 BPM, whisper-quiet with a very low spectral centroid. Cavern rumble and slow water drips; a heavily reverberated gasp opens the moment and an ominous swell lands on the reveal.

To match the suffocating darkness of this cave, we need an audio landscape that feels incredibly vast yet intimately claustrophobic, anchored by an ultra-low-energy, ambient C minor drone that immediately establishes a sense of dread. Operating at a whisper-quiet level with a very low spectral centroid, the mix must favor deep, dark frequencies, like distant cavern rumbles and the cold, slow dripping of water, to mirror the flickering shadows and the agonizingly slow pan of the camera. The underlying musical tempo sits at a patient one hundred and eighteen beats per minute, but it should feel almost imperceptible, serving as a slow, tense pulse rather than a driving rhythm to keep the audience suspended in anticipation. While there is no formal dialogue, that initial, fragile breath at the very beginning of the timeline must be heavily reverberated, sounding like a sharp gasp lost in the cavernous space to signal the protagonist's immediate vulnerability. Finally, as the glowing eyes pierce the darkness, the spatial audio must shift, pulling the low-frequency drone tighter around the listener while a subtle, ominous swell synchronizes with the visual reveal to solidify the presence of this ancient, lurking danger.

Emotion

ominous. Across scene 20’s eight moments: mysterious, suspenseful, calm, sorrowful, eerie, tense, ominous, tragic, sad.

Synthetic sentiment

On the synthetic curve below, scene 20 reads +1.8 for all viewers, +2.3 for ages 18–24 and +1.0 for ages 25–34, predicted before anyone watched.

Search tags

dragon reveal · creature horror · cave · dread · unseen threat

And more

moment ID · thumbnail · character presence · segment-level sentiment · technical audio metrics

What the film told Arc View

The film tells a richer story than its catalog record. The record says fantasy, Millie Bobby Brown, princess, dragon. Arc View traced six storylines, from Elodie’s resistance to a forced marriage to the dragon curse rotting the kingdom, and each one became a searchable thread.

Six storylines detected and tracked throughout the timeline.

One film, two records

a catalog record

genre: Fantasy
cast: Millie Bobby Brown
tags: princess · dragon

the Arc View record

structure: 37 scenes, 193 moments
storylines: [6] from Elodie’s resistance to the dragon curse
characters: Elodie 30.2%, Prince Henry 5.6%, Lady Bayford 5.1%, Floria 3.9% of runtime
every moment described: narrative, video technical, audio technical, story beat, emotion tags
synthetic sentiment: moment-to-moment sentiment simulated per audience cohort
traditional tags: genre, emotion and more

Elodie, sitting among flames on the cave floor, faces the dragon in Damsel's final scene.

Moments are where search gets specific. The eye in the cave wall could be exactly what someone means when they ask for a dragon movie that actually scares them, or when they want to find a scene they enjoyed to show a friend.

Allowing people to relive specific moments effortlessly.

See film sentiment before you put it in front of any audience

Every story borrows from a short list. Our thesis: the world holds 13 characters and 7 stories, and everything on screen is a new arrangement of them. Break a film down to the moment and the story beat, and its pieces match beats audiences have already watched.

13 characters. 7 stories. Every film rearranges them.

Pilotly has captured over 30 million real audience reaction data points, moment by moment. When a new title arrives, Arc View compares each of its moments to similar story beats in that data and carries over the reactions viewers had to those moments. Combined across the whole film, those reactions form a synthetic sentiment curve: a prediction of how each audience segment is likely to feel from start to finish before anyone has watched it.

Synthetic sentiment across Damsel, before any audience saw it

36:26 The eye in the cave: the dragon reveal from the moment viewer above.

Synthetic sentiment across Damsel, before any audience saw it.
Time	All viewers	Ages 18–24	Ages 25–34
0:50	+1.5	+1.6	+1.7
2:30	+1.8	+2.5	+1.8
4:20	+1.8	+2.4	+1.5
6:00	+1.8	+2.0	+1.9
8:40	+2.2	+2.1	+2.4
12:10	+1.9	+2.1	+1.6
14:50	+1.9	+1.8	+1.6
16:30	+1.7	+1.9	+1.0
18:40	+2.0	+2.1	+1.8
21:10	+2.4	+3.0	+1.9
23:40	+1.2	+1.3	+0.7
25:30	+2.1	+2.5	+1.3
28:40	+2.2	+2.0	+1.8
33:50	+2.0	+1.9	+1.4
38:00	+1.8	+2.3	+1.0
44:00	+1.8	+2.5	+1.2
52:20	+1.8	+2.6	+1.0
57:40	+2.0	+2.5	+1.4
61:50	+2.0	+3.0	+1.7
65:50	+1.9	+2.8	+1.2
68:50	+2.0	+2.9	+1.6
73:40	+1.8	+2.5	+1.4
77:40	+1.9	+2.1	+1.8
79:20	+2.0	+2.3	+1.9
81:30	+1.8	+2.3	+1.0
83:40	+1.8	+3.1	+1.3
86:20	+2.0	+2.4	+1.9
89:00	+2.1	+2.8	+2.0
93:30	+2.2	+2.6	+2.0
98:40	+2.6	+3.2	+2.6
101:20	+2.9	+3.4	+2.5

Arc View synthetic sentiment. Read from the dashboard chart to about ±0.1, times to about ±30 seconds; the end-of-film drop to zero is left out.

For Damsel, that curve ranks the moments most likely to pull a member in, segment by segment. Search can lead with those moments, and recommendations can match the film to the members its scenes are most likely to move.

What we took to Netflix

A repeatable process by which they could generate a language-model-ready taxonomy for their content. Generated from the content with an interface where they can keep humans in the loop for spot checking, now search could match the story a member describes to various titles in their library.

193 moments, each one searchable.

A bridge to the data Netflix already has. Every description maps to a timecode, so the taxonomy joins existing catalog metadata, viewing data and artwork tests at the level of the scene.

Story metadata that joins to viewing data.

A moment for every search, and the feeling of the moment that can match to member preference. Synthetic sentiment ranks Damsel’s moments by likely appeal for each audience segment, so a result can open on the scene most likely to pull that member in, and recommendations can surface titles whose scenes move members the way they have responded before.

The right ten seconds for the right member.

Damsel was the proof of concept for Netflix. The same pipeline can now run across a title perpetually and turn a catalog into something a language model can interpret to drive new experience value for platform members.

More case studies

The Mandalorian & Grogu

film — march 2026

long episodic — december 2025

Mars Reel x Pizza Hut

branded content — april 2019

Bring us the next one.

We'll design the study around the decision riding on it.

book a demo tell us what project you're working on