Add TSFM feature catalog store

INSTRUCTIONS.md

@@ -85,17 +85,18 @@ See [MCP Servers](#mcp-servers) for available tools and [docs/mcp-servers.md](do
 
 ## Environment Variables
 
-**CouchDB** — `iot` and `wo` servers
-
-| Variable           | Default                 | Description              |
-| ------------------ | ----------------------- | ------------------------ |
-| `COUCHDB_URL`      | `http://localhost:5984` | CouchDB connection URL   |
-| `COUCHDB_USERNAME` | `admin`                 | CouchDB admin username   |
-| `COUCHDB_PASSWORD` | `password`              | CouchDB admin password   |
-| `IOT_DBNAME`         | `iot`                   | IoT sensor database name      |
-| `WO_DBNAME`          | `workorder`             | Work order database name      |
-| `FAILURE_CODE_DBNAME` | `failure_code`          | FCC failure-code database name |
-| `VIBRATION_DBNAME`   | `vibration`             | Vibration sensor database name |
+**Database-backed servers** — shared database connection and per-server database names
+
+| Variable                   | Default                 | Description                        |
+| -------------------------- | ----------------------- | ---------------------------------- |
+| `COUCHDB_URL`              | `http://localhost:5984` | CouchDB connection URL             |
+| `COUCHDB_USERNAME`         | `admin`                 | CouchDB admin username             |
+| `COUCHDB_PASSWORD`         | `password`              | CouchDB admin password             |
+| `IOT_DBNAME`               | `iot`                   | IoT sensor database name           |
+| `WO_DBNAME`                | `workorder`             | Work order database name           |
+| `FAILURE_CODE_DBNAME`      | `failure_code`          | FCC failure-code database name     |
+| `VIBRATION_DBNAME`         | `vibration`             | Vibration sensor database name     |
+| `FEATURE_CATALOG_DBNAME`   | `feature_catalog`       | TSFM feature catalog database name |
 
 **WatsonX** — plan-execute runner and WatsonX-backed agent routes
 
@@ -137,15 +138,15 @@ See [MCP Servers](#mcp-servers) for available tools and [docs/mcp-servers.md](do
 
 ## MCP Servers
 
-Six FastMCP servers cover IoT data, time-series ML, work orders, vibration diagnostics, failure-mode reasoning, and utility tools. They speak MCP over stdio and are spawned on-demand by the agent runners — no manual startup needed.
+Six FastMCP servers cover IoT data, time-series feature catalogs, work orders, vibration diagnostics, failure-mode reasoning, and utility tools. They speak MCP over stdio and are spawned on-demand by the agent runners — no manual startup needed.
 
 | Server      | Tools | Categories               | Backing service                        |
 | ----------- | ----- | ------------------------ | -------------------------------------- |
 | `iot`       | 7     | read                     | CouchDB  (telemetry + asset registry)  |
 | `utilities` | 3     | read                     | none                                   |
 | `fmsr`      | 2     | read, LLM-use            | LiteLLM + `failure_modes.yaml`         |
 | `wo`        | 15    | read, write              | CouchDB                                |
-| `tsfm`      | 6     | read, write, cpu-centric | IBM Granite TinyTimeMixer (torch)      |
+| `tsfm`      | 8     | read, write              | feature catalog database               |
 | `vibration` | 8     | read, cpu-centric        | CouchDB + numpy/scipy DSP              |
 
 Tool signatures, required env vars, and how to launch a server directly: **[docs/mcp-servers.md](docs/mcp-servers.md)**.
@@ -413,7 +414,7 @@ Each integration suite is gated by a `skipif` mark; missing service ⇒ silently
 | ------------------ | ---------------------------------------------------------------------------- |
 | iot, wo, vibration | CouchDB reachable — `docker compose -f src/couchdb/docker-compose.yaml up -d` |
 | fmsr               | `WATSONX_APIKEY`, `WATSONX_PROJECT_ID` set in `.env`                          |
-| tsfm               | `PATH_TO_MODELS_DIR`, `PATH_TO_DATASETS_DIR` set in `.env`                    |
+| tsfm               | Feature catalog database loaded, or `TSFM_STORE=memory` for hermetic tests    |
 
 Narrow scope by path or name pattern:
 

README.md

@@ -86,7 +86,7 @@ AssetOpsBench is a **unified framework for developing, orchestrating, and evalua
 |---|---|
 | **IoT** | `sites`, `asset_ids`, `asset_detail`, `assets`, `find_assets_by_sensors`, `installed_sensors`, `measured_sensors` |
 | **FMSR** | `get_failure_modes`, `generate_failure_modes`, `add_failure_modes`, `generate_failure_mode_sensor_mapping` |
-| **TSFM** | `forecasting`, `timeseries_anomaly_detection` |
+| **TSFM** | `list_features`, `search_features`, `get_feature`, `register_feature`, `new_feature_version` |
 | **WO** | `get_work_order_distribution`, `predict_next_work_order`, ... |
 | **Vibration** | `compute_fft_spectrum`, `compute_envelope_spectrum`, ... |
 | **...** | **...** |
@@ -112,7 +112,7 @@ The `src/` directory contains MCP servers and a plan-execute runner built on the
 |---|---|
 | **IoT** | "List all assets in MAIN site" |
 | **FMSR** | "List known failure modes for asset class pump" |
-| **TSFM** | "Forecast Chiller 9 Condenser Water Flow for the week of 2020-04-27" |
+| **TSFM** | "Find active time-series normalization features in the feature catalog" |
 | **WO** | "Generate a work order for Chiller 6 anomaly detection" |
 
 Some tasks focus on a single domain, others are multi-step end-to-end workflows. Explore all scenarios on [Hugging Face](https://huggingface.co/datasets/ibm-research/AssetOpsBench).

docs/mcp-servers.md

@@ -8,7 +8,7 @@ Six FastMCP servers expose the AssetOpsBench domain logic. Each is a standalone
 - [utilities — Utilities](#utilities--utilities)
 - [fmsr — Failure Mode and Sensor Relations](#fmsr--failure-mode-and-sensor-relations)
 - [wo — Work Order](#wo--work-order)
-- [tsfm — Time Series Foundation Model](#tsfm--time-series-foundation-model)
+- [tsfm — Time Series Feature Catalog](#tsfm--time-series-feature-catalog)
 - [vibration — Vibration Diagnostics](#vibration--vibration-diagnostics)
 
 ## iot — IoT Asset Registry and Telemetry Records
@@ -109,20 +109,32 @@ _None — the WO server makes no LLM calls; all tools are direct CouchDB operati
 
 _None — all tools are lightweight CouchDB queries/mutations (Mango `_find` / `GET` / `PUT`), with no heavy computation._
 
-## tsfm — Time Series Foundation Model
+## tsfm — Time Series Feature Catalog
 
 **Path:** `src/servers/tsfm/main.py`
-**Requires:** `tsfm_public` (IBM Granite TSFM), `transformers`, `torch` for ML tools — imported lazily; static tools work without them.
-**Model checkpoints:** resolved relative to `PATH_TO_MODELS_DIR` (default: `src/servers/tsfm/artifacts/output/tuned_models`)
-
-| Tool                   | Category                 | Arguments                                                                                                                   | Description                                                                                      |
-| ---------------------- | ------------------------ | --------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------ |
-| `get_ai_tasks`         | read                     | —                                                                                                                           | List supported AI task types for time-series analysis                                            |
-| `get_tsfm_models`      | read                     | —                                                                                                                           | List available pre-trained TinyTimeMixer (TTM) model checkpoints                                 |
-| `run_tsfm_forecasting` | read, write, cpu-centric | `dataset_path`, `timestamp_column`, `target_columns`, `model_checkpoint?`, `forecast_horizon?`, `frequency_sampling?`, ...  | Zero-shot TTM inference; returns path to a JSON predictions file                                 |
-| `run_tsfm_finetuning`  | read, write, cpu-centric | `dataset_path`, `timestamp_column`, `target_columns`, `model_checkpoint?`, `save_model_dir?`, `n_finetune?`, `n_test?`, ... | Few-shot fine-tune a TTM model; returns saved checkpoint path and metrics file                   |
-| `run_tsad`             | read, write, cpu-centric | `dataset_path`, `tsfm_output_json`, `timestamp_column`, `target_columns`, `task?`, `false_alarm?`, `ad_model_type?`, ...    | Conformal anomaly detection on top of a forecasting output JSON; returns CSV with anomaly labels |
-| `run_integrated_tsad`  | read, write, cpu-centric | `dataset_path`, `timestamp_column`, `target_columns`, `model_checkpoint?`, `false_alarm?`, `n_calibration?`, ...            | End-to-end forecasting + anomaly detection in one call; returns combined CSV                     |
+**Requires:** configured catalog database; `FEATURE_CATALOG_DBNAME` selects the database name (default: `feature_catalog`).
+**Catalog data:** `src/couchdb/scenarios_data/shared/tsfm/feature_catalog.json`
+
+The TSFM server manages feature catalog cards. Transform cards store executable
+EFE-style `fit` / `transform` programs; extractor cards store searchable metadata
+for scalar feature extractors.
+
+| Tool | Category | Arguments | Description |
+| ---- | -------- | --------- | ----------- |
+| `list_features` | read | `kind?`, `status?` | List transform and/or extractor cards. `kind` may be `transform`, `extractor`, or omitted. |
+| `search_features` | read | `text?`, `tags?`, `status?` | Search cards by feature id, name, description, or tags. |
+| `get_feature` | read | `feature_id` | Return one stored feature card by id. |
+| `register_feature` | write | `feature`, `overwrite?` | Register a transform card after schema and executable-code validation. |
+| `update_feature` | write | `feature_id`, `fields` | Patch metadata fields on an existing card without rerunning executable validation. |
+| `deprecate_feature` | write | `feature_id`, `reason?` | Mark a card deprecated while keeping it available for audit and lineage. |
+| `new_feature_version` | write | `feature_id`, `fields?`, `new_feature_id?` | Create a validated successor transform card and mark the predecessor superseded. |
+| `get_feature_lineage` | read | `feature_id` | Return parent and direct-descendant ids for a feature card. |
+
+Successful TSFM tool responses include a top-level `message` string. List and
+search responses also include `features`; registration returns `status`, `id`,
+and `card`; card operations return the card fields plus `message`; lineage
+returns `feature_id`, `ancestors`, `root`, `descendants`, and `message`. Errors
+return `ErrorResult` with an `error` field.
 
 ## vibration — Vibration Diagnostics
 

pyproject.toml

@@ -31,8 +31,6 @@ dependencies = [
     "stirrup[docker,litellm,mcp]>=0.1.7",
     "python-dotenv>=1.0",
     "scipy>=1.10.0",
-    "sktime>=1.0.1",
-    "granite-tsfm>=0.3.5",
 ]
 
 [project.scripts]
@@ -59,12 +57,6 @@ dev = [
     "opentelemetry-api>=1.27.0",
     "opentelemetry-sdk>=1.27.0",
 ]
-# Optional heavy ML deps for the TSFM server.
-# tsfm_public must be installed separately: pip install git+https://github.com/ibm-granite/granite-tsfm
-tsfm = [
-    "torch>=2.0",
-    "transformers>=4.40",
-]
 # Optional OpenTelemetry tracing for agent runners.
 # Enable by setting OTEL_EXPORTER_OTLP_ENDPOINT (e.g. http://localhost:4318).
 otel = [
@@ -84,4 +76,3 @@ anyio_mode = "auto"
 filterwarnings = [
     "ignore:Core Pydantic V1 functionality:UserWarning",
 ]
-