Larkup Video Intelligence

An installable Marketplace v3 tool with a cloud-first Larkup Cloud connection. On first use, the app creates an opaque, device-scoped key automatically; administrators may also provide a replacement key in Installed Tools. The AWS GPU endpoint and raw key are never shown in the connection status. PyAV/FFmpeg, faster-whisper, RapidOCR/PaddleOCR, YOLOX, ONNX Runtime, and anonymous tracking are isolated in runtime images.

The evidence pipeline is genre-neutral. The chat model turns the meaning of a question into a typed investigation directive: local, temporal, or whole-source scope plus an answer, comparison, trace, enumeration, or synthesis operation. The executor validates and performs that operation without selecting a workflow from question words, language, or video genre. Direct visual claims preserve generic subject, relation, and value bindings so the same corroboration logic works across every source. Captioning progresses in chronological waves: only identities grounded by source evidence become continuity candidates for a later wave, and visual resemblance alone never establishes identity.

Where things live

runtime/app/
  main.py          FastAPI app + CORS + exception handlers, mounts the v1 router
  config.py        Settings -- the one place env vars are read
  api/
    v1.py          every /v1 route (health, uploads, jobs, usage, access codes)
    deps.py        auth/rate-limit dependencies, and the settings/store/jobs singletons
  services/
    brain.py       bounded agent planner: chooses modalities, density, segments, and extraction focus
    pipeline.py    shared cloud/local executor: scout -> plan -> parallel evidence extraction -> synthesis
    transcription.py  TranscriptionService: WhisperProvider (local) / DeepgramProvider (hosted)
    vision.py      SemanticVisionService: per-clip captioning via the Vercel AI Gateway
    embedding.py   VideoEmbeddingService: cross-modal clip search via DashScope (Qwen3-VL-Embedding)
    scene.py       SceneDetector: scene-cut + fixed-window clip planning
    motion.py      MotionSampler: adaptive frame selection within a clip
    storage.py     pluggable frame-upload interface (S3 today; not hardcoded to it)
    jobs.py        JobService: runs a queued job in a background thread (local-Docker path only)
  db/
    store.py       SQLite-backed principals/keys/uploads/jobs
    schemas.py     pydantic request/response models
  utils/
    timing.py      range clamping/merging, sampling cadence, timestamp rebasing
  worker.py        local-Docker CLI entrypoint (`python -m app.worker <job_id>`)

deploy/ is the operator/infrastructure side for Larkup's managed cloud offering: the AWS control plane (Lambda + API Gateway + DynamoDB + S3) and GPU provider dispatch/worker entrypoints (RunPod, Modal, and other vendors behind one interface). It is public for review and collaboration, but remains fully separate from runtime/app so a self-hosted Docker/local install never depends on it. It uses short-lived CI identity and contains no deployed credentials. See deploy/README.md for testing, validation, and deploy commands.

Current roadmap

Collaborators are currently focused on three connected improvements:

1. Smarter indexing — produce a coherent story-like timeline and durable notes that preserve the important people, questions, events, and changes across the full video.
2. Faster analysis — reduce indexing and follow-up latency with bounded parallel work, better sampling, and direct inspection of only the ranges that need more evidence.
3. Stronger reuse of indexing results — make analysis and chat use the stored transcript, visual evidence, entities, summaries, and answering guide first, then request bounded re-analysis only when indexed evidence is incomplete.

Local run

Requirements: Docker Desktop or Docker Engine with Compose v2. NVIDIA acceleration additionally requires the NVIDIA Container Toolkit.

pnpm --filter @larkup/tool-video-intelligence runtime:start
# NVIDIA GPU (CUDA + PaddleOCR):
pnpm --filter @larkup/tool-video-intelligence runtime:start -- --gpu
curl http://127.0.0.1:8787/v1/health

Published-package users run the same lifecycle command:

npx @larkup/tool-video-intelligence start

The local default builds the cross-platform CPU target with RapidOCR, binds only port 8787, does not require an account, and keeps sources/state in Docker volumes. The GPU override uses the CUDA target with PaddleOCR. Set LARKUP_VIDEO_REQUIRE_AUTH=true and LARKUP_VIDEO_ADMIN_TOKEN when exposing a self-hosted endpoint. Do not expose the unauthenticated local configuration to a public network.

Runtime configuration

The package-level .env.example is the single configuration surface for local Docker plus optional cloud GPU providers. It is passed into Compose as the container environment, so a value changed there takes effect on the next runtime restart. Never commit the generated .env.

# Create packages/marketplace-tools/video-intelligence/.env once.
pnpm --filter @larkup/tool-video-intelligence exec larkup-video-intelligence config init

# Update a credential, request rate, or GPU selection.
pnpm --filter @larkup/tool-video-intelligence exec larkup-video-intelligence config set AI_GATEWAY_API_KEY ...
pnpm --filter @larkup/tool-video-intelligence exec larkup-video-intelligence config set LARKUP_VIDEO_REQUESTS_PER_MINUTE 240
pnpm --filter @larkup/tool-video-intelligence exec larkup-video-intelligence config set LARKUP_VIDEO_ACCELERATOR gpu
pnpm --filter @larkup/tool-video-intelligence runtime:start

config get KEY reads a configured value and config path prints the exact file location. The CLI only accepts documented keys from .env.example, which prevents a misspelled credential or GPU variable from silently doing nothing.

Runtime choices in Larkup

The installed-tool settings card is driven by this tool's manifest and offers three compatible runtimes:

• Larkup Cloud is the default. Paste a Larkup Cloud API key in Installed Tools when your organization issues keys manually, or leave it empty when automatic project provisioning is enabled. The same card has Request Larkup Cloud API key for requesting access; monthly usage is shown only for this metered mode.
• Local runtime detects Docker and uses it when available; otherwise it installs and runs the native CPU runtime with uv. The card shows the chosen engine, host suitability, install state, and explicit Install, Start, and Stop controls. It generates a shared API key and lets you replace 127.0.0.1 with a private LAN URL for trusted users.
• Custom runtime accepts a compatible /v1 endpoint and bearer key. Use the built-in Verify action before indexing through a self-hosted provider.

Every runtime requires an audio provider key plus agent/tool-brain and vision credentials. The audio model is selected automatically for the provider. Agent and vision controls start with the provider, model, and key saved under AI Models, and each can be overridden independently in the tool settings. A text-only provider such as DeepSeek can remain the agent while vision uses a separate provider. Local mode injects these settings only into the local process/container. Managed Cloud forwards them only in the active GPU job and never stores them in DynamoDB or S3. Larkup Cloud supplies compute, not model credits. The agent/tool-brain model may differ from the vision model.

For a Docker-free portable runtime, use the uv package in runtime/:

# Install uv once: https://docs.astral.sh/uv/
npx @larkup/tool-video-intelligence native

The Installed Tools card exposes this automatically selected native path as part of Local runtime. It creates ./.larkup/video-intelligence/ for state and downloaded models, binds to 127.0.0.1 by default, and uses the same /v1 API and SQLite-backed worker queue as Docker. To use a LAN URL, set the card's local URL to an address on the host and keep the generated key private. Configure optional AI/provider credentials with larkup-video-intelligence config set … before starting.

Indexing contract

1. Upload to POST /v1/uploads.
2. Submit POST /v1/jobs with a typed indexing brief.
3. Poll GET /v1/jobs/{id}.
4. Index the returned timestamped transcript, OCR, object tracks, and answering guide.

The only indexing modes are fast, balanced, and thorough. They are bounded latency/recall budgets, not fixed pipelines. A first planning call chooses useful modalities, a model-free scout collects sparse chronological motion and optional OCR signals while transcription runs in parallel, and a second planning call chooses the final whole-source cadence plus denser source-supported priority ranges. Every proposed number is clamped by the executor. The same Python implementation runs locally and in managed cloud.

Progress reports measured frame/clip counters, never moves backwards, and adds a remaining-time estimate calculated from the actual plan. Results include the plan, model-call diagnostics, elapsed time, estimate error, and timestamped evidence so a caller can audit both speed and extraction quality.

For managed cloud jobs, the Data-library asset is the canonical copy. A short-lived, KMS-encrypted processing copy is created only so the GPU worker can decode it; it is deleted when the job settles, including failed and cancelled jobs. Once the app has downloaded the evidence, it acknowledges the result and the cloud result object is deleted too. S3 lifecycle rules remove any abandoned processing source or result within one day as a safety bound.

For a complex chat question, the agent routes the question before answering: focused facts use ranked timestamped evidence, a request for every source-authored question or written item uses an unranked chronological scan, and people/timeline requests can use a revision-scoped aggregate retained outside the chat context. The source mapper separates content-bearing prompts from interactional speech and declarative clues, stores directly established grids, sequences, and collections independently from their readable prompt text, and records separately resolved instances of a reusable prompt when the source proves their boundaries. That lets chat distinguish a complete source scan from a partially legible layout instead of treating banter or hints as questions. A full-source result is marked complete only when every source-inventory window finished; otherwise the response stays explicitly partial. Answers and respondents attached to an inventoried question are retained only when nearby source evidence contains them. If the selected evidence is incomplete, the agent may request one authorized, bounded cloud re-analysis (maximum 30 seconds per request) and then reads the resulting evidence again. The worker seeks directly to the approved range for model inference; source transfer and scale-to-zero GPU startup can still add latency for large videos.

Answer memory and corrections

Chat keeps a durable answer memory for exact repeated video questions. A result is saved only after the active source evidence supports it, and is scoped to that video's active knowledge revision. Reindexing or a bounded inspection creates a new revision, so an older cached answer is never reused against new evidence.

Unsupported answers are not cached. If the same unanswered question is asked again, the app records the repeat and may perform a bounded reinspection of the best candidate range instead of returning the earlier "unknown" result.

When a user explicitly corrects a video answer in chat, the assistant saves that correction as an opt-in, user-confirmed answer for exact repeats. It is kept separate from source evidence: it can make the next identical question fast, but it never rewrites the indexed transcript, OCR, or visual evidence and does not manufacture a citation.

Managed-cloud entitlement

The runtime does not know about Stripe or another payment provider. API keys carry a provider-neutral entitlement:

{
  "plan": "access-code",
  "sourceMinutesPerMonth": 600,
  "maxConcurrentJobs": 2
}

Monthly source-minute reservations and concurrency updates are atomic. A future subscription webhook only needs to create/update this entitlement; video processing does not change.

The dashboard warns at 80% usage and can send a temporary capacity request through Resend. Configure these server-only environment variables in apps/web to enable delivery:

RESEND_API_KEY=...
LARKUP_USAGE_REQUEST_FROM='Larkup <[email protected]>'
LARKUP_USAGE_REQUEST_TO='[email protected]'

For managed jobs, the audio, brain, and vision credentials travel through the authenticated control plane only as a transient worker payload. They are not included in the persisted job item or result, and the worker restores a clean environment after completion. Provider failures do not fall back to a Larkup-funded model.

For a controlled pilot, set AutoProvisioningEnabled=true. The first TrialDeviceLimit new device IDs (100 by default) receive TrialSourceMinutes (600 minutes, or 10 hours, by default) each month; later devices receive PostTrialSourceMinutes (0 by default) and can use the dashboard's support request instead. This is an allowance and rate-limit system only—there is no customer-facing price or payment flow.

The support team can update a device's allowance, concurrency, or per-minute request limit without seeing its API key:

curl -X POST "$VIDEO_ENDPOINT/v1/admin/devices/$USER_ID/entitlement" \
  -H "Content-Type: application/json" \
  -H "X-Larkup-Admin-Token: $VIDEO_ADMIN_TOKEN" \
  -d '{"sourceMinutesPerMonth":60,"maxConcurrentJobs":1,"requestsPerMinute":60,"plan":"support-grant"}'

For the common “add credits” workflow, use the support script. It calls the authenticated control plane, which updates the DynamoDB device entitlement and reconciles the current billing-period allowance atomically; it does not write to a Neon database. The minutes value is the new total monthly allowance, not an increment.

export LARKUP_VIDEO_INTELLIGENCE_CLOUD_ENDPOINT='https://video.example.com'
export LARKUP_VIDEO_ADMIN_TOKEN='...'
bash scripts/grant-cloud-credits.sh <user-id> 120 support-grant

$USER_ID is the generated ID shown in Installed Tools and included in the Resend support request. Each local Larkup installation receives a separate key, while DynamoDB stores only hashes of keys and installation identifiers plus aggregate usage.

Access codes remain available for support-issued or upgraded plans:

curl -X POST "$VIDEO_ENDPOINT/v1/admin/access-codes" \
  -H "Content-Type: application/json" \
  -H "X-Larkup-Admin-Token: $VIDEO_ADMIN_TOKEN" \
  -d '{"label":"pilot","sourceMinutesPerMonth":600,"maxConcurrentJobs":2,"maxUses":1}'

Only hashes of access codes, API keys, and installation identifiers are stored. The raw device key is saved only in the local project's configuration so it can authenticate future cloud jobs.

AWS deployment

The production stack -- API Gateway + Lambda control plane, DynamoDB, S3 + KMS, and GPU dispatch through Modal/RunPod -- is real-deployed and smoke-tested end to end (control plane → GPU worker → S3 → job completion). Its source lives in the public deploy/ directory described above. ProcessingEnabled defaults to false -- a stack accepts no real jobs until explicitly turned on.

See deploy/README.md for the build/deploy, bootstrap, and RunPod image publish commands. Do not send a long-lived AWS secret through chat or store it in GitHub; a short AWS SSO session is sufficient for bootstrap, with GitHub assuming a narrow OIDC deployment role afterward.

Operator-level env vars

These are self-hosted worker/service variables. Managed Cloud receives the corresponding user values per job and has no platform-owned model-provider secrets (see tool.manifest.json's configSchema):

| Var | Default | Purpose | | ---------------------------------------------------- | ---------------------------- | ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | | LARKUP_VIDEO_VISION_PROVIDER | vercel_ai_gateway | vercel_ai_gateway, google, or openai; defaults from AI Models and can be overridden per tool. | | LARKUP_VIDEO_VISION_API_KEY | unset | User-owned selected provider key, supplied from AI Models or a per-tool override. | | LARKUP_VIDEO_SEMANTIC_VISION_MODEL | google/gemini-3.6-flash | Bulk per-clip captioning model. Gemini is called through Google's native API when LARKUP_VIDEO_VISION_PROVIDER=google. | | LARKUP_VIDEO_AGENT_PROVIDER | vercel_ai_gateway | Defaults from the AI Models chat provider and can be overridden per tool; supports Gateway, Google, OpenAI, DeepSeek, Mistral, Cohere, and Anthropic. | | LARKUP_VIDEO_AGENT_API_KEY | unset | User-owned agent provider credential. | | LARKUP_VIDEO_AGENT_MODEL | openai/gpt-5-mini | Tool-brain model that chooses modalities, sampling density, priority ranges, and extraction focus. | | LARKUP_VIDEO_CONTINUITY_WAVE_BATCHES | 4 | Parallel vision batches per chronological wave. Only source-grounded continuity candidates cross from one wave to the next. | | LARKUP_VIDEO_GOOGLE_CONCURRENCY | 4 | Maximum simultaneous native Gemini vision batches. Kept below Gateway concurrency to avoid direct-project burst limits. | | LARKUP_VIDEO_GOOGLE_MAX_IMAGES_PER_REQUEST | 8 | Bounds native Gemini request size so multi-clip structured responses remain within interactive provider timeouts. | | LARKUP_VIDEO_GOOGLE_REQUESTS_PER_MINUTE | 12 | Sliding-window request limit for native Gemini vision calls, leaving quota headroom for the agent model. | | LARKUP_VIDEO_REASONING_VISION_MODEL | google/gemini-3.6-flash | Dense final-verification model for bounded source inspection. | | LARKUP_VIDEO_TRANSCRIPTION_PROVIDER | unset | User-selected timestamped speech provider: deepgram, openai, groq, or elevenlabs. | | LARKUP_VIDEO_TRANSCRIPTION_FALLBACK | unset | Managed Cloud keeps this empty so it never falls back to Larkup-funded inference. | | LARKUP_VIDEO_TRANSCRIPTION_CHUNK_SECONDS | 180 | Audio window used for long hosted transcriptions. Windows keep request sizes bounded and are merged back onto the original source clock. | | LARKUP_VIDEO_TRANSCRIPTION_REQUEST_TIMEOUT_SECONDS | 60 | Maximum wait for one hosted speech request before the configured fallback takes over, preventing a stalled provider from holding the progress UX for minutes. | | LARKUP_VIDEO_DEEPGRAM_AUTO_MODEL | nova-3 | User-selected Deepgram model used when the source language is unknown. | | LARKUP_VIDEO_TRANSCRIPTION_CONCURRENCY | 3 | Maximum parallel hosted speech windows. The runtime keeps the merged transcript chronological even when requests finish out of order. | | LARKUP_VIDEO_GPU_PROVIDER | modal | modal or runpod; the managed-cloud GPU dispatch registry, private (see deploy/README.md). | | LARKUP_VIDEO_EMBEDDING_PROVIDER | gateway-gemini-embedding-2 | gateway-gemini-embedding-2 uses the configured AI Gateway key and Gemini's multimodal vector space for image-frame/text retrieval. disabled, DashScope qwen3-vl-embedding, runpod-qwen3-vl-embedding, and huggingface-qwen3-vl-embedding remain available. Live bounded answer verification skips embeddings entirely because vectors are not needed to return timestamped answer evidence. | | LARKUP_VIDEO_EMBEDDING_FALLBACK_PROVIDER | gateway-gemini-embedding-2 | Provider used automatically when the selected visual embedding service is unavailable. Set disabled to require the primary provider only. | | DASHSCOPE_API_KEY | unset | Required when LARKUP_VIDEO_EMBEDDING_PROVIDER=qwen3-vl-embedding. | | DASHSCOPE_WORKSPACE_ID | unset | From the Model Studio console's Workspace Details page. Required when LARKUP_VIDEO_EMBEDDING_PROVIDER=qwen3-vl-embedding. | | DASHSCOPE_REGION | unset | The workspace's region, e.g. eu-central-1 (same console page). Required when LARKUP_VIDEO_EMBEDDING_PROVIDER=qwen3-vl-embedding. | | LARKUP_VIDEO_RUNPOD_EMBEDDING_ENDPOINT_ID | unset | Required when LARKUP_VIDEO_EMBEDDING_PROVIDER=runpod-qwen3-vl-embedding (also needs RUNPOD_API_KEY). | | LARKUP_VIDEO_HF_EMBEDDING_URL | unset | Required when LARKUP_VIDEO_EMBEDDING_PROVIDER=huggingface-qwen3-vl-embedding (also needs HF_TOKEN). | | LARKUP_MEDIA_STORAGE | local | Web app env var: s3 switches createStorageProvider() to S3StorageProvider for canonical media (see LARKUP_MEDIA_S3_BUCKET etc.). | | WEBHOOK_SIGNING_SECRET | unset | Signs outbound job-completion webhooks (X-Larkup-Signature). | | STALE_JOB_TIMEOUT_HOURS | 6 | Force-fails a queued/running job older than this on the scheduled reconcile sweep. |

Verification

pnpm --filter @larkup/tool-video-intelligence type-check
pnpm --filter @larkup/tool-video-intelligence test
pnpm --filter @larkup/marketplace test   # S3StorageProvider, tool-loader, etc.
PYTHONPATH=runtime python3 -m unittest discover -s runtime/tests -p 'test_*.py'
docker build --target smoke -t larkup-video-smoke runtime
docker build --target cpu -t larkup-video-cpu runtime

The AWS control plane and GPU provider tests/validation run privately -- see deploy/README.md.

The runtime is the default path for indexed videos. Removing this package, its v3 catalog record, and the optional installed-tool entry cleanly removes the implementation.