A Case in Real Footage: When Search Becomes Editing
OpenMontage's documentary-montage pipeline tests whether its agent-first design can coordinate a video made from real footage rather than generated stills. The key question is not whether the system can download clips; it is how it turns a thematic brief into a searchable corpus, then turns ranked candidates into an edited argument without confusing similarity with meaning. The case gives the reader a concrete way to judge retrieval, provenance, motion, and human editorial choice together.
This path matters because “video from open sources” can mean two very different things. It can mean collecting a few images and moving a camera across them, or it can mean finding actual motion clips, choosing useful moments, and arranging them into a timeline. OpenMontage explicitly treats documentary montage as the second problem. The pipeline is labeled beta, so the case demonstrates a supported mechanism and its boundaries, not a guarantee that every topic will yield a finished film.
The corpus is a search index, not a shelf of clips
The documentary pipeline begins with a thematic brief and scene slots. A slot is not just a file request such as “find a city clip.” It describes the visual subject, mood, scale, era, or narrative role that the edit needs. The scene director supplies short search queries and preferred sources, while the asset stage chooses whether to use a standard corpus-and-retrieval path or a faster direct-search path.
The standard path builds a local corpus by fanning queries across available sources such as stock libraries, open archives, and institutional collections. The corpus builder downloads candidates, records their source information, extracts video thumbnails, probes dimensions and duration, measures a simple motion signal, and computes CLIP embeddings. It also stores a tag or description channel so a candidate can be represented both by what its frames look like and by the source's textual metadata. The result is an appendable search index, not an undifferentiated folder.
That design changes how the agent works. It does not browse files in the order they were downloaded and accept the first plausible result. It queries the corpus for each slot. A rich slot description becomes the text used for semantic ranking; short stock-search queries are used to gather candidates, but they are not necessarily the best representation of what the final shot must communicate.
The faster path makes a different trade-off. Direct clip search fans out across available providers, downloads a small number of candidates per query, extracts thumbnails, and expects a person or agent to inspect them. It avoids the cost and time of building an embedding index, which is useful for act-by-act production or a smaller number of slots. The two paths can coexist: direct search can fill most of a montage quickly, while corpus retrieval can address a difficult slot whose visual match remains weak.
The standard path also has a useful operational property: it can grow rather than restart. Existing clips remain in the corpus, later queries can add candidates, and a failed or partial candidate does not have to invalidate every other row. The corpus builder records errors per source or candidate and can reuse work already downloaded. That makes retrieval a persistent production resource, not a one-time web search that disappears after the first edit. It also makes the quality question more concrete: if the corpus is too small, too static, or too concentrated in one source, the agent has evidence for expanding or reshaping it.