Skip to content
Fluvie Fluvie Fluvie

Give AI evidence about your assets

Prepare a local catalog before asking a model to tell a story. The catalog binds media facts, selected observations and transcripts to the source file’s SHA-256. Contact sheets record actual source frames and presentation times. Creating a catalog does not call an authoring provider or invent visual descriptions.

Terminal window
fluvie assets ./assets --contact-sheets \
  --catalog-out build/fluvie/assets.catalog.json
fluvie generate "the life of my cat" \
  --catalog build/fluvie/assets.catalog.json --context-file assets/story.txt \
  --dart-out lib/cat_story.dart --no-render

The catalog and sheets remain in your project. Authoring verifies their source and image hashes before using them; replacing an asset requires rebuilding its catalog. A stale observation must not silently describe a different video.

Each sheet records its selected frames, source times, layout and toolchain provenance. Its cache key covers those inputs and the source hash; generated filenames are an implementation detail. Read the sheet path and cacheKey from the catalog rather than deriving a filename from the video name or source hash.

Supply observations with their certainty

Section titled “Supply observations with their certainty”

Write a JSON array of selected observations. Asset names are exact paths relative to the selected assets directory; times are source seconds. Keep facts separate from your interpretation through observed, inferred or unknown certainty.

[
  {
    "asset": "cat/playing.mp4",
    "fromSeconds": 1,
    "toSeconds": 2.5,
    "text": "The cat jumps onto the sofa.",
    "certainty": "observed"
  }
]

Save this as assets/observations.json and select it explicitly:

Terminal window
fluvie assets ./assets --catalog-out build/fluvie/assets.catalog.json \
  --contact-sheets --evidence observations.json

Observations record the selected file’s provenance. Text alone does not imply that the model viewed the footage. Uncertain cues should remain uncertain in the authored story.

Bind timestamped transcripts

Section titled “Bind timestamped transcripts”

Use an SRT or WebVTT file that already exists; Fluvie does not transcribe audio automatically. Both sides of --transcript are relative to the assets directory unless the caption path is absolute. Repeat it for several sources:

Terminal window
fluvie assets ./assets --catalog-out build/fluvie/assets.catalog.json \
  --transcript cat/playing.mp4=cat/playing.vtt

The catalog preserves source-second ranges, caption text and provenance. Selected transcript files and observations are bounded; excessively large inputs fail with an actionable message rather than entering an unlimited prompt.

Let a vision provider inspect selected sheets

Section titled “Let a vision provider inspect selected sheets”

Contact-sheet creation is local. Uploading them to your selected provider requires --image-evidence on authoring:

Terminal window
fluvie generate "the life of my cat" \
  --catalog build/fluvie/assets.catalog.json --image-evidence \
  --context-file assets/story.txt --no-render

Authoring selects at most four verified sheets, each up to 4 MiB. Use a vision-capable provider and select an image-capable --model when needed; Ollama’s default llama3.1 model is text-only. Adapters preserve selected cues through their image transport, combining labeled sheets for a single-image provider API when necessary. These cues remain available during repair attempts; the model should not claim to see footage when only metadata or text was supplied. Your provider receives the selected pictures. Use your own local inspection and explicit observations when you want to keep pictures on your machine.

Where to next

Section titled “Where to next”