Database Model

This document describes where and in what form the AI stack stores data. There are four repositories, each with a distinct role. For operational procedures, see Database Administration.

Most are not owned by AI

PostgreSQL and Cassandra are shared with the PlantPulse Platform installation. The AI stack does not run its own databases. This means "AI backup" and "Platform backup" overlap — if you don't track who owns what, coverage gaps appear on both sides.

Repository	Owner	Contents
PostgreSQL `ch`	AI (separate DB within Platform PG)	Chat · incidents · settings · MCP
Cassandra `pi`	AI	Feature store · anomaly detection outputs
Cassandra `pp`	Platform (read-only)	Raw sensor · alarm · OEE
Neo4j	AI	Ontology knowledge graph
Qdrant	AI	Document embedding vectors

PostgreSQL `ch` — Conversation and operational state

A separate database ch within the Platform PostgreSQL instance (jdbc:postgresql://<platform-host>:5432/ch).

Group	Tables
Conversation	`chats` · `chat_messages` · `prompts` · `prompt_history`
Users	`users` · `user_roles` · `employees` · `web_push_subscription`
Incidents	`incident` · `incident_event` · `assignees`
Anomaly detection operations	`anomaly_mute` · `hourly_anomaly` · `asset_tag_profile`
MCP	`mcp_server` · `mcp_tool_call_log` · `mcp_tool_pref`
RAG integration	`rag_asset_card` · `rag_asset_doc_sync` · `rag_sync_run`
Settings	`config` · `config_audit_log` · `metadata`

This is the single source of truth for chat history

Before the July 2026 reorganization, chat history also lived in Cassandra. Now it exists only here — chat_messages is the authoritative source. Do not look in Cassandra based on older documentation.

Cassandra `pi` — Feature store

Twelve tables owned and created by the time series insights service. Prefix is tsi_.

Group	Tables
Features	`tsi_tag_feature_hourly` · `tsi_tag_feature_daily` · `tsi_asset_feature_hourly` · `tsi_asset_feature_daily`
Context · profile	`tsi_asset_context_daily` · `tsi_asset_profile`
Anomaly · fault	`tsi_anomaly_embedding` · `tsi_failure_signature` · `tsi_precursor_alert`
Pipeline	`tsi_pipeline_status` · `tsi_pipeline_history` · `tsi_backfill_status`

Tables are created idempotently when the service starts. However, the keyspace pi itself is not created by the application — it is managed by the installation script (provision-platform-db.sh).

Empty features blank out the entire analytics view

The admin console's anomaly detection and forecasting screens read entirely from these tables. If results are empty, check the pipeline first, not the model → admin console Pipeline Status · Run History (AI Admin Console).

The old pp.ti_* does not exist

On 2026-07-30, the pp keyspace was migrated to pi and the old tables were dropped. Queries using the old name will return no results.

Cassandra `pp` — Platform source (read-only)

AI reads only from Platform-collected data. It does not write here.

Table	Purpose
`tm_asset_data_based_second` / `_minute` / `_hour`	Equipment sensor time series
`tm_tag_point`	Tag data points
`tm_asset_alarm`	Alarms
`tm_asset_oee_history`	OEE
`tm_asset_ram` · `tm_asset_ems_history`	RAM · EMS
`tm_monitor_platform_status`	System resources (CPU · memory · disk)

Partition key differs by resolution

second : PRIMARY KEY ((asset_id, date, hour), timestamp)
minute : PRIMARY KEY ((asset_id, date), timestamp)
hour   : PRIMARY KEY (asset_id, timestamp)

If queries are slow, check resolution first

Second resolution includes equipment · date · time in the partition. Scanning long periods at second resolution touches thousands of partitions. Use _hour for wide ranges, _second only for narrow ranges.

The complete Platform schema is in Platform Database Model.

Neo4j · Qdrant — Knowledge graph and vectors

Repository	Contents
Neo4j	Ontology — relationship graph of equipment · processes · documents
Qdrant	Document chunk embedding vectors

The RAG engine (LightRAG) is configured to store graphs in Neo4j and vectors in Qdrant (LIGHTRAG_GRAPH_STORAGE=Neo4JStorage, LIGHTRAG_VECTOR_STORAGE=QdrantVectorDBStorage).

Recovery requires re-ingesting documents

Both are derived from source documents. Without backups, recovery means re-uploading the original files and rebuilding the indexes — keeping the original files is effectively your backup.

For ontology concepts, see Understanding Ontology.

Backup coverage checklist

Repository	Backed up by	Loss impact
PostgreSQL `ch`	Must be handled on AI side	Chat · incidents · settings disappear
Cassandra `pi`	Must be handled on AI side	Features disappear, analytics screens go blank (recalculation possible but slow)
Cassandra `pp`	Platform backup	—
Neo4j · Qdrant	AI	Document indexes disappear (can be regenerated from source)

For procedures, see Database Administration.

Database Administration — access · inspection · operations
AI Admin Console — pipeline status
Service Port Map
Platform Database Model

PostgreSQL ch — Conversation and operational state​

Cassandra pi — Feature store​

Cassandra pp — Platform source (read-only)​

Partition key differs by resolution​

Neo4j · Qdrant — Knowledge graph and vectors​

Backup coverage checklist​

Related documents​