Local-First Video AI: What Stays Local and What Uses the Cloud
Evaluate local-first video AI by tracing original media, derived data, indexes, queries, retention, and connectivity instead of relying on a single privacy label.
Local-first video AI keeps authoritative source media under the user's control, but the label does not automatically mean that every frame, embedding, query, or model operation stays offline. The useful question is not “local or cloud?” It is: what data crosses each boundary, for what purpose, under which retention and access rules?
For ShotAI's current documented workflow, original video files remain on the user's computer. The desktop app reads local media, prepares thumbnails, detects shots, stores library information and search vectors, and reconnects results to source files for playback and export. AI understanding and semantic indexing can process derived visual information—and, when present, derived audio information—in cloud services. Semantic queries are also encoded online before local vector retrieval. See the current privacy and connectivity documentation before processing restricted material.
What “Local-First” Should Mean
A defensible local-first design starts with local ownership and useful offline state, then documents every remote dependency. It should answer five questions:
- Where do the original files live?
- Which derivatives are created, such as frames, thumbnails, audio features, transcripts, or embeddings?
- Which derivatives or queries leave the device?
- Where are indexes, metadata, and source references stored?
- What stops working when the network or account service is unavailable?
“Original files are not uploaded” is meaningful, but it is not equivalent to “nothing about the footage leaves the device.” Derived frames, audio features, transcripts, prompts, and telemetry can still reveal sensitive content. Procurement and security reviews should trace them separately.
ShotAI's Current Processing Boundary
The current product documentation describes this split:
| Workflow component | Current documented location or behavior | Review question |
|---|---|---|
| Original video files | Remain on the user's computer | Are source paths and storage permissions controlled? |
| Playback and export | Access local source files | What happens if a drive moves or disconnects? |
| Thumbnails and shot detection | Prepared by the desktop workflow | Which generated artifacts are retained locally? |
| Library information and search vectors | Stored for local library retrieval | How are indexes backed up, deleted, and protected? |
| AI understanding and semantic indexing | Uses cloud processing of derived visual and possible audio data | What is transmitted, retained, logged, or used for service operation? |
| Semantic query encoding | Query text is processed online | Can the query itself reveal confidential project details? |
This boundary is narrower and more accurate than describing ShotAI as fully offline. When semantic services are unavailable, the application may fall back to exact search; local AI functions do not replace the documented cloud indexing and semantic-query workflow.
Local-First Is Not the Same as Air-Gapped
Local-first software may use online sign-in, licensing, model services, updates, or query processing. Air-gapped software operates without a network path to external systems. Private deployment is a separate contractual and technical claim that must be verified for a specific version and agreement.
Do not infer air-gapped operation, a retention period, a training-use policy, a storage region, or a compliance certification from the phrase “local-first.” Ask the vendor for current documentation and applicable terms.
When the Architecture Matters Most
The processing boundary deserves extra review for:
- unreleased productions and client work under confidentiality terms;
- interviews, documentary sources, and footage containing personal data;
- legal, healthcare, education, government, or regulated-industry material;
- archives whose size makes full-media upload slow or costly;
- teams that must keep masters on controlled storage while adding a discovery layer.
The decision remains policy-specific. Keeping originals local may satisfy one requirement while cloud processing of derived data violates another. Only the applicable agreement and security review can decide.
A Practical Vendor Evaluation Checklist
Use representative, permitted footage and record the answers:
- Originals: Are full files uploaded, copied, modified, or cached elsewhere?
- Derivatives: Are frames, audio, transcripts, embeddings, or proxies transmitted?
- Queries: Where are natural-language searches encoded and logged?
- Retention: What is stored, for how long, and how is deletion verified?
- Access: Which employees, subprocessors, or connected assistants can receive data?
- Offline behavior: Which import, exact search, semantic search, playback, and export actions still work?
- Source reconnection: Can every result return to the correct file and time range after drives move?
- Governance: Are permissions, rights, approvals, backups, and audit logs handled here or in a DAM/MAM?
For broader procurement, use the video asset management buyer's guide. For retrieval behavior, review semantic video search.
Product Boundary
ShotAI is a discovery layer for a user's own indexed footage. It connects natural-language, shot-level results to local sources and editing workflows. It is not a preservation master, rights authority, complete multi-user DAM/MAM, or proof that material is safe to process. Keep authoritative rights, consent, retention, and backup controls in the systems responsible for them.
FAQ
Does ShotAI upload original video files?
The current product documentation says original files remain on the user's computer. Derived visual data, possible audio-derived data, and semantic query text can use cloud services.
Does local-first mean fully offline?
No. Local-first and air-gapped are different. Verify every online dependency and the behavior when it is unavailable.
Is derived data harmless because it is smaller than the original?
No. A frame, transcript, embedding, or query can reveal sensitive content even when the full source file is never uploaded.
How should a team validate the claim?
Trace originals, derivatives, indexes, queries, telemetry, retention, access, deletion, and offline behavior using current documentation and a permitted test library.