Why MongoDB

The video models can be swapped. The value is in the data they produce, and MongoDB stores, links, searches and delivers that data on one platform.

One platform for the data the models produce

Operational documents, time series, geospatial, full-text and vector search on one platform, with one security model and no sync pipelines. Every box, card and chart in this demo opens the MongoDB document or query behind it.

The agent's memory is in the same database

LangGraph's short-term memory (MongoDBSaver checkpoints) and long-term memory (MongoDBStore) run on MongoDB next to the incidents they refer to: no separate memory store or vector database. What the agent learns is governed like any other data: TTL expiry, access control, backup and audit.

Search and vector indexes stay in sync

The false positive disappeared from the next search without a manual reindex: search and vector indexes sync automatically from the database, so operators never see stale results from drift.

Atlas today, Enterprise Advanced tomorrow

Built on Atlas in Singapore; the same code and queries run on Enterprise Advanced 8.3+ with Search and Vector Search in GCC or on premises. Only the connection string changes. No Triggers, Functions, Stream Processing or automated embedding in the runtime path.

Reliable delivery without a broker

Incident and outbox entry in one transaction; adapters resume from change-stream tokens; a unique index blocks duplicates. We killed the adapter mid-flight: nothing lost, nothing duplicated.

The video models can be swapped

Video stays in the VMS; MongoDB holds metadata, vectors and media URIs with SHA-256. Every generated document carries modelVer, so models change without changing the platform.

Scale

~83 M

VLM vectors over 30 days at full scale

340 GB → 85 GB

vector storage, float32 → int8 quantization

4.8 TB → 1.35 TB

total storage when indexing events only

7,016

cameras on the island-wide map

This demo's footprint: 60 corridor cameras with full data, 90,000 detections over 30 days (50 per camera per day) with 1,024-dimension Voyage embeddings and int8-quantized vector indexes, VLM observations every 15 minutes for the last 7 days (production rate: every 30 seconds), 60 seeded long-term memories. Server …: $rankFusion fallback, vector quantization off.

All incidents, cameras, plates and operators are SIMULATED. Demonstration only. Not affiliated with or endorsed by LTA.