@larkup/tool-video-intelligence 0.2.0

package/.env.example

@@ -0,0 +1,150 @@
+# This is the one configuration file for a self-hosted Video Intelligence
+# runtime and its optional GPU providers. Copy it with:
+#   pnpm --filter @larkup/tool-video-intelligence exec larkup-video-intelligence config init
+# Then update individual values without opening an editor:
+#   pnpm --filter @larkup/tool-video-intelligence exec larkup-video-intelligence config set KEY VALUE
+
+# Docker/runtime selection. Set LARKUP_VIDEO_ACCELERATOR=gpu to make the
+# lifecycle command use compose.gpu.yaml; `--gpu` remains a one-off override.
+LARKUP_VIDEO_IMAGE=ghcr.io/larkup-ai/video-intelligence:0.1.0
+LARKUP_VIDEO_PORT=8787
+LARKUP_VIDEO_ACCELERATOR=cpu
+LARKUP_VIDEO_DEVICE=auto
+LARKUP_VIDEO_WORKERS=1
+LARKUP_VIDEO_DATA_DIR=/data
+LARKUP_VIDEO_MODEL_DIR=/models
+LARKUP_VIDEO_ALLOWED_ORIGINS=http://localhost:3000
+LARKUP_VIDEO_MAX_UPLOAD_BYTES=21474836480
+LARKUP_VIDEO_REQUESTS_PER_MINUTE=120
+LARKUP_VIDEO_REQUIRE_AUTH=false
+LARKUP_VIDEO_SHARED_API_KEY=
+LARKUP_VIDEO_ADMIN_TOKEN=
+
+# Vercel AI Gateway powers vision and the recommended multimodal embedding
+# provider. It can route OpenAI models too, so one gateway credential is all
+# this runtime needs for those model calls.
+AI_GATEWAY_API_KEY=
+AI_GATEWAY_BASE_URL=https://ai-gateway.vercel.sh/v1
+LARKUP_VIDEO_SEMANTIC_VISION=true
+LARKUP_VIDEO_SEMANTIC_VISION_MODEL=google/gemini-3.6-flash
+LARKUP_VIDEO_REASONING_VISION_MODEL=google/gemini-3.6-flash
+LARKUP_VIDEO_SEMANTIC_REASONING_EFFORT=minimal
+# Tool-brain planner. Every runtime receives the user's selected provider,
+# model, and key; managed cloud supplies GPU compute only.
+LARKUP_VIDEO_AGENT_ENABLED=true
+LARKUP_VIDEO_AGENT_PROVIDER=vercel_ai_gateway
+LARKUP_VIDEO_AGENT_MODEL=openai/gpt-5-mini
+LARKUP_VIDEO_AGENT_API_KEY=
+LARKUP_VIDEO_AGENT_BASE_URL=https://ai-gateway.vercel.sh/v1
+LARKUP_VIDEO_GATEWAY_BATCH_SIZE=4
+LARKUP_VIDEO_GATEWAY_CONCURRENCY=24
+LARKUP_VIDEO_GATEWAY_REQUESTS_PER_MINUTE=60
+LARKUP_VIDEO_GOOGLE_CONCURRENCY=4
+LARKUP_VIDEO_GOOGLE_MAX_IMAGES_PER_REQUEST=8
+LARKUP_VIDEO_GOOGLE_REQUESTS_PER_MINUTE=12
+LARKUP_VIDEO_GATEWAY_EMBEDDING_MODEL=google/gemini-embedding-2
+LARKUP_VIDEO_GATEWAY_EMBEDDING_BATCH_SIZE=6
+LARKUP_VIDEO_GATEWAY_EMBEDDING_CONCURRENCY=6
+LARKUP_VIDEO_GATEWAY_EMBEDDING_BASE_URL=https://ai-gateway.vercel.sh/v4/ai
+
+# User-owned hosted speech-to-text. Managed cloud never injects a provider key
+# and does not fall back to a Larkup-funded local Whisper pass.
+LARKUP_VIDEO_TRANSCRIPTION_PROVIDER=deepgram
+LARKUP_VIDEO_TRANSCRIPTION_FALLBACK=
+# Long hosted transcriptions are split once and sent in bounded parallel requests.
+LARKUP_VIDEO_TRANSCRIPTION_CHUNK_SECONDS=180
+LARKUP_VIDEO_TRANSCRIPTION_CONCURRENCY=3
+LARKUP_VIDEO_TRANSCRIPTION_REQUEST_TIMEOUT_SECONDS=60
+LARKUP_VIDEO_WHISPER_MODEL=small
+DEEPGRAM_API_KEY=
+LARKUP_VIDEO_DEEPGRAM_MODEL=nova-3
+LARKUP_VIDEO_DEEPGRAM_AUTO_MODEL=nova-3
+OPENAI_API_KEY=
+LARKUP_VIDEO_OPENAI_TRANSCRIPTION_MODEL=whisper-1
+GROQ_API_KEY=
+LARKUP_VIDEO_GROQ_TRANSCRIPTION_MODEL=whisper-large-v3-turbo
+ELEVENLABS_API_KEY=
+LARKUP_VIDEO_ELEVENLABS_TRANSCRIPTION_MODEL=scribe_v2
+
+# Cross-modal video-clip embeddings. The gateway provider is the default;
+# choose one of: disabled, gateway-gemini-embedding-2, qwen3-vl-embedding,
+# runpod-qwen3-vl-embedding, huggingface-qwen3-vl-embedding.
+LARKUP_VIDEO_EMBEDDING_PROVIDER=gateway-gemini-embedding-2
+# If the selected service is unavailable, keep RAG retrieval operational with
+# the user-owned AI Gateway key. Set to disabled to require the primary only.
+LARKUP_VIDEO_EMBEDDING_FALLBACK_PROVIDER=gateway-gemini-embedding-2
+LARKUP_VIDEO_EMBEDDING_MODEL=qwen3-vl-embedding
+LARKUP_VIDEO_EMBEDDING_DIMENSION=3072
+LARKUP_VIDEO_EMBEDDING_FRAMES_PER_CLIP=4
+LARKUP_VIDEO_EMBEDDING_INSTRUCTION=
+DASHSCOPE_API_KEY=
+DASHSCOPE_WORKSPACE_ID=
+DASHSCOPE_REGION=
+LARKUP_VIDEO_DASHSCOPE_BASE_URL=
+RUNPOD_API_KEY=
+LARKUP_VIDEO_RUNPOD_EMBEDDING_API_KEY=
+LARKUP_VIDEO_RUNPOD_EMBEDDING_ENDPOINT_ID=
+LARKUP_VIDEO_RUNPOD_EMBEDDING_BASE_URL=
+HF_TOKEN=
+LARKUP_VIDEO_HF_EMBEDDING_API_KEY=
+LARKUP_VIDEO_HF_EMBEDDING_URL=
+
+# Frame object storage. Keep local data URIs by default; S3 needs the normal
+# AWS credential chain supplied by the deployment environment.
+LARKUP_VIDEO_FRAME_STORAGE=local
+LARKUP_VIDEO_BUCKET=
+LARKUP_VIDEO_FRAME_PREFIX=tmp-frames
+AWS_REGION=eu-central-1
+LARKUP_VIDEO_OCR_ENGINE=RapidOCR
+LARKUP_VIDEO_DISABLE_HEAVY_OPERATORS=false
+
+# Cloud control-plane and GPU-provider settings. They are inert for local
+# Docker use, but are intentionally kept here so a deployment has no second
+# hidden env file.
+LARKUP_VIDEO_GPU_PROVIDER=modal
+LARKUP_VIDEO_GPU_TYPE=
+RUNPOD_ENDPOINT_ID=
+RUNPOD_REGISTRY_AUTH_ID=
+MODAL_TOKEN_ID=
+MODAL_TOKEN_SECRET=
+MODAL_API_KEY=
+LARKUP_VIDEO_MODAL_APP=larkup-video-intelligence
+LARKUP_VIDEO_MODAL_FUNCTION=process_video_job
+GPUAI_API_KEY=
+HYPERSTACK_API_KEY=
+HYPERSTACK_ENVIRONMENT_NAME=
+HYPERSTACK_KEY_NAME=
+NORTHFLANK_API_KEY=
+NORTHFLANK_PROJECT_ID=
+NORTHFLANK_DEPLOYMENT_PLAN=
+NORTHFLANK_IMAGE_CREDENTIALS_ID=
+SALAD_API_KEY=
+SALAD_ORGANIZATION_NAME=
+SALAD_PROJECT_NAME=
+SCALEAWAY_ACCESS_KEY_ID=
+SCALEAWAY_API_KEY=
+SCALEAWAY_DEFAULT_ZONE=
+SCALEAWAY_GPU_IMAGE_ID=
+SCALEAWAY_PROJECT_ID=
+SHADE_FORM_API_KEY=
+THUNDERCOMPUTE_API_KEY=
+VAST_API_KEY=
+
+# AWS control-plane policy and webhook configuration.
+ADMIN_TOKEN=
+WEBHOOK_SIGNING_SECRET=
+MAX_UPLOAD_BYTES=21474836480
+REQUESTS_PER_MINUTE=120
+PROCESSING_ENABLED=false
+AUTO_PROVISIONING_ENABLED=false
+# The first 100 new device IDs receive a 600-minute (10-hour) monthly cloud
+# allowance. No payment provider is involved; later devices can request support access.
+TRIAL_DEVICE_LIMIT=100
+TRIAL_SOURCE_MINUTES=600
+TRIAL_REQUESTS_PER_MINUTE=60
+POST_TRIAL_SOURCE_MINUTES=0
+# Backward-compatible allowance for cloud device records created before the
+# bounded trial policy. New devices use the TRIAL_* values above.
+AUTO_PROVISIONED_SOURCE_MINUTES=30
+STALE_JOB_TIMEOUT_HOURS=6
+TESTING_DEVICE_HASHES=

package/LICENSE

@@ -0,0 +1,176 @@
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS

package/README.md

@@ -0,0 +1,281 @@
+# 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. Planning is driven by the user's goal,
+question shape, source duration, and measured visual/audio signals—not a fixed
+branch for sports, lectures, entertainment, films, meetings, or any other
+content type. Direct visual claims preserve generic `subject`, `relation`, and
+`value` bindings so the same corroboration logic works across every genre.
+
+## 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 gitignored and not part of this public checkout -- kept fully separate from `runtime/app` so a self-hosted Docker/local install never depends on it. If you operate the managed cloud stack, see `deploy/README.md` in your private checkout for testing, validation, and deploy commands.
+
+## Local run
+
+Requirements: Docker Desktop or Docker Engine with Compose v2. NVIDIA acceleration additionally requires the NVIDIA Container Toolkit.
+
+```bash
+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:
+
+```bash
+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`](./.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`.
+
+```bash
+# 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/`:
+
+```bash
+# 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 first searches active timestamped evidence. If it is incomplete, it may request one authorized, bounded cloud re-analysis (maximum 30 seconds per request) to add fresh OCR, visual detections, or anonymous tracks, then searches the resulting evidence again. This avoids treating RAG snippets as the final answer. 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:
+
+```json
+{
+  "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:
+
+```bash
+RESEND_API_KEY=...
+LARKUP_USAGE_REQUEST_FROM='Larkup <usage@example.com>'
+LARKUP_USAGE_REQUEST_TO='support@example.com'
+```
+
+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:
+
+```bash
+curl -X POST "$VIDEO_ENDPOINT/v1/admin/devices/$USER_ID/entitlement" \
+  -H "Content-Type: application/json" \
+  -H "X-Larkup-Admin-Token: $VIDEO_ADMIN_TOKEN" \
+  -d '{"sourceMinutesPerMonth":60,"maxConcurrentJobs":1,"requestsPerMinute":60,"plan":"support-grant"}'
+```
+
+For the common “add credits” workflow, use the support script. It calls the
+authenticated control plane, which updates the DynamoDB device entitlement and
+reconciles the current billing-period allowance atomically; it does not write
+to a Neon database. The minutes value is the new total monthly allowance, not
+an increment.
+
+```bash
+export LARKUP_VIDEO_INTELLIGENCE_CLOUD_ENDPOINT='https://video.example.com'
+export LARKUP_VIDEO_ADMIN_TOKEN='...'
+bash scripts/grant-cloud-credits.sh <user-id> 120 support-grant
+```
+
+`$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:
+
+```bash
+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 private, gitignored `deploy/` directory described above, not in this public checkout. `ProcessingEnabled` defaults to `false` -- a stack accepts no real jobs until explicitly turned on.
+
+See `deploy/README.md` in a private checkout 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_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
+
+```bash
+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.

package/compose.gpu.yaml

@@ -0,0 +1,14 @@
+services:
+  video-intelligence:
+    build:
+      target: runtime
+    environment:
+      LARKUP_VIDEO_DEVICE: cuda
+      LARKUP_VIDEO_OCR_ENGINE: PaddleOCR
+    deploy:
+      resources:
+        reservations:
+          devices:
+            - driver: nvidia
+              count: 1
+              capabilities: [gpu]