diff --git a/CHANGELOG.md b/CHANGELOG.md
index dd24e0b..4a26a8c 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -4,6 +4,19 @@
 - The format is based on [Keep a Changelog](https://keepachangelog.com/).
 - This project adheres to [Semantic Versioning](https://semver.org/).
 
+
+## Version 1.1.0 - Upcoming
+
+### Added
+
+- New `@UI.RecommendationState` opt-in annotation for scalar fields without a value help (free-form numerics, free-text). Annotated fields are added to the entity's `<Entity>_Recommendations` companion alongside value-helped fields. RPT-1 `task_type` is selected per column: numeric scalars opted in via `@UI.RecommendationState` use `regression` so the model can interpolate continuous values; everything else uses `classification` (existing behaviour). The `task_type` enum on `predictRowColumns` is widened from `{classification}` to `{classification, regression}`.
+
+### Fixed
+- CDS-to-RPT-1 dtype map now matches the inference API's enum (`'string' | 'numeric' | 'date'`). Previously emitted `'bool'` for `cds.Boolean` and `'datetime'` for `cds.DateTime` / `cds.Timestamp`, causing HTTP 422 from `/predict` for any entity carrying those types. `cds.DateTime` and `cds.Timestamp` now map to `'string'` so the full ISO value is preserved as an opaque token (no time loss, no date-parse rejection).
+- Composition children of draft-enabled entities are now reliably enhanced. The CSN enhancer previously only walked `entity.compositions` — a legacy CSN shape — and missed the `entity.elements[*]` form used by current CDS, with the result that nested entities (e.g. `Approvers` under a `ChangeRequests` draft) never received `@UI.Recommendations` and so never got recommendations even when their fields had value lists. The walk is now recursive through both shapes with cycle protection.
+- Empty-rows crash in `_fetchPrediction`. Reading a draft entity whose composition was empty fed `rows: []` to the prediction API, and the schema-derivation reduce dereferenced `rows[0][ele]` on `undefined`, taking the whole server down on a TypeError. `_fetchPrediction` now returns an empty result for empty input. The READ handler also short-circuits when the response set is empty, avoiding a needless AI Core round-trip.
+- RPT-1 inference limits now honoured. When `target_columns.length > 10` or when a row carries more than 100 columns, `_fetchPrediction` logs a warning and returns an empty result instead of letting the API reject the request with a noisy 422. The warning includes a hint to use `@UI.RecommendationState : 0` to opt fields out and bring the count down.
+
 ## Version 1.0.1 - 2026-05-08
 
 ### Fixed
diff --git a/README.md b/README.md
index 2bd5311..5ff2be2 100644
--- a/README.md
+++ b/README.md
@@ -58,6 +58,53 @@ annotate Books with {
 }
 ```
 
+#### Recommendations on fields without a value help
+
+By default the plugin only enhances fields that have a value list — that's how it auto-detects which columns are good prediction targets. Some fields are good targets but have no value list: free-form numerics like measurement ranges, calibration values, or planning estimates. Annotate these with `@UI.RecommendationState : 1` to opt in:
+
+```cds
+entity CalibrationData : cuid {
+  measuringRangeMin : Decimal(16, 6) @UI.RecommendationState : 1;
+  measuringRangeMax : Decimal(16, 6) @UI.RecommendationState : 1;
+  operatingPoint    : Decimal(16, 6) @UI.RecommendationState : 1;
+  description       : String         @UI.RecommendationState : 1;
+}
+```
+
+The annotation only takes effect on **scalar** elements (no associations / compositions / unmanaged elements; for those, attach a value help instead). Annotated fields are added to the entity's `<Entity>_Recommendations` companion just like value-helped fields, and Fiori Elements' soft-fill placeholder renders the prediction in the empty input.
+
+`task_type` is chosen automatically per column:
+- numeric scalar (`Integer*`, `Decimal`, `Double`) annotated with `@UI.RecommendationState` → **`regression`** so RPT-1 can interpolate continuous values,
+- everything else → **`classification`**.
+
+> [!NOTE]
+> Numeric fields that have a value help (e.g. a fixed price-point list) stay on classification — `@UI.RecommendationState` is only needed when there is *no* value help. Combining both is unnecessary.
+
+#### How recommendations work under the hood
+
+A short FAQ for integrators, so you don't have to read the source.
+
+**What does the plugin emit on the OData service?**
+On every draft-enabled entity that has at least one value-helped field, it adds an entity-level annotation `@UI.Recommendations: { '=': 'SAP_Recommendations' }` plus a synthetic companion entity (`<Entity>_Recommendations`, `@cds.persistence.skip`) with one virtual array per recommendable field. Each item carries `RecommendedFieldValue`, `RecommendedFieldDescription`, `RecommendedFieldScoreValue` and `RecommendedFieldIsSuggestion` — the shape Fiori Elements expects for `UI.RecommendationListType`. The first entry per field has `RecommendedFieldIsSuggestion: true` and is rendered as the soft-fill default.
+
+**When does it run?**
+On READ requests to a draft entity that expand `SAP_Recommendations`. Reads against the active entity return nothing in that field. Reads during `draftActivate` are skipped.
+
+**What data is sent to RPT-1 as context?**
+Up to 2000 rows from the **active** version of the same entity, restricted to rows where every recommendable field is non-null. The columns `createdAt`, `createdBy`, `modifiedAt`, `modifiedBy` plus any `cds.LargeBinary` / `cds.Vector` elements are stripped. The active row corresponding to the draft (if any) is removed and replaced by the draft row carrying `[PREDICT]` placeholders in the columns to predict. There is no sampling or `ORDER BY` — for tables larger than 2000 rows, which rows make the cut is determined by the database.
+
+> [!IMPORTANT]
+> Everything in the remaining columns is forwarded to AI Core. Annotate sensitive fields with `@UI.RecommendationState : 0` (or a dynamic expression) to keep them out of both the predictions and the context payload.
+
+**How are descriptions populated?**
+For each predicted value, the plugin issues an extra SELECT against the field's `@Common.Text` association (if set) to fetch the human-readable label. Fields without `@Common.Text` get an empty `RecommendedFieldDescription`.
+
+**RPT-1 deployment lifecycle**
+First prediction call against a resource group provisions an `sap-rpt-1-small` deployment in scenario `foundation-models` (executable `aicore-sap`) and polls up to 10× with exponential backoff until it reaches `RUNNING`. Subsequent calls reuse the cached deployment. Single-tenant uses the configured `resourceGroup` (default `'default'`); multi-tenant creates one resource group per tenant on subscribe (label `ext.ai.sap.com/CDS_TENANT_ID`) and deletes it on unsubscribe.
+
+**Local development**
+Without an AI Core binding the plugin uses `MockAICoreService`, which returns the first non-null value of each target column from the context as the "prediction" — useful for UI smoke tests, useless as a quality signal. Run `cds bind <your-aicore-instance>` and start with profile `hybrid` to talk to a real AI Core deployment locally.
+
 ### 2. Use case: Simplified AI Core usage
 
 The plugin introduces an `AICore` CAP service that automatically performs some administrative tasks and offers simplified access to AI Core.
diff --git a/lib/csn-enhancements/recommendations.js b/lib/csn-enhancements/recommendations.js
index 5912b8a..cc8867e 100644
--- a/lib/csn-enhancements/recommendations.js
+++ b/lib/csn-enhancements/recommendations.js
@@ -9,11 +9,39 @@ function enhanceModel(model) {
       continue;
     }
     enhanceEntity(name, model);
-    if (entity.compositions) {
-      for (const comp in entity.compositions) {
-        enhanceEntity(entity.compositions[comp].target, model);
+    // Walk composition children recursively. In legacy CSN flavours the
+    // compositions are exposed under `entity.compositions`; in current
+    // (CDS >= 7) CSN they are part of `entity.elements` with
+    // `type === 'cds.Composition'`. enhanceEntity itself is idempotent
+    // (it bails if `@UI.Recommendations` is already set), so duplicate
+    // visits are safe; the visited set is just a micro-optimisation and
+    // a guard against composition cycles.
+    const visited = new Set([name]);
+    const walk = (entityName) => {
+      const e = model.definitions[entityName];
+      if (!e) return;
+      if (e.compositions) {
+        for (const comp in e.compositions) {
+          const tgt = e.compositions[comp].target;
+          if (tgt && !visited.has(tgt)) {
+            visited.add(tgt);
+            enhanceEntity(tgt, model);
+            walk(tgt);
+          }
+        }
       }
-    }
+      if (e.elements) {
+        for (const ele in e.elements) {
+          const def = e.elements[ele];
+          if (def.type === 'cds.Composition' && def.target && !visited.has(def.target)) {
+            visited.add(def.target);
+            enhanceEntity(def.target, model);
+            walk(def.target);
+          }
+        }
+      }
+    };
+    walk(name);
   }
   model.meta[enhancedFlag] = true;
 }
@@ -27,22 +55,34 @@ function enhanceEntity(name, model) {
   const entity = model.definitions[name];
   if (entity['@UI.Recommendations']) return; // already enhanced
   const vhFields = Object.keys(entity.elements).reduce((vhFields, ele) => {
-    // check if the property has a value help
+    const def = entity.elements[ele];
+    if (def['@UI.RecommendationState'] === 0) return vhFields;
+    // 1) Auto-detected: property has a value help.
     const hasValueHelp =
-      entity.elements[ele]['@Common.ValueList.CollectionPath'] ||
-      model.definitions[entity.elements[ele].target]?.['@cds.odata.valuelist'];
-    if (entity.elements[ele]['@UI.RecommendationState'] !== 0 && hasValueHelp) {
-      if (entity.elements[ele].keys) {
-        for (const key of entity.elements[ele].keys) {
+      def['@Common.ValueList.CollectionPath'] ||
+      model.definitions[def.target]?.['@cds.odata.valuelist'];
+    if (hasValueHelp) {
+      if (def.keys) {
+        for (const key of def.keys) {
           vhFields[ele + '_' + key.ref.join('_')] = structuredClone(
-            model.definitions[entity.elements[ele].target].elements[key.ref]
+            model.definitions[def.target].elements[key.ref]
           );
           delete vhFields[ele + '_' + key.ref.join('_')].key;
         }
-      } else if (!entity.elements[ele].on) {
-        vhFields[ele] = structuredClone(entity.elements[ele]);
+      } else if (!def.on) {
+        vhFields[ele] = structuredClone(def);
         delete vhFields[ele].key;
       }
+      return vhFields;
+    }
+    // 2) Opt-in: scalar field annotated `@UI.RecommendationState` without a value
+    // help. Lets RPT-1 predict free-form values (typically numeric) for
+    // fields where Fiori Elements' soft-fill placeholder is the desired
+    // UX. Skips associations / compositions / unmanaged elements — those
+    // are handled by the value-help branch above.
+    if (def['@UI.RecommendationState'] && !def.target && !def.on) {
+      vhFields[ele] = structuredClone(def);
+      delete vhFields[ele].key;
     }
     return vhFields;
   }, {});
diff --git a/lib/handlers/recommendations.js b/lib/handlers/recommendations.js
index 6a14728..79c690a 100644
--- a/lib/handlers/recommendations.js
+++ b/lib/handlers/recommendations.js
@@ -96,6 +96,10 @@ export default function registerHandlersForRecommendations(srv) {
       if (!res) return res;
       const fieldsWithDisabledRecommendations = {};
       const response = Array.isArray(res) ? res : [res];
+      // No subjects means nothing to predict for. Returning early avoids an
+      // empty-rows round-trip to AI Core (which would also be a degenerate
+      // input for RPT-1).
+      if (response.length === 0) return res;
       const aiCore = await cds.connect.to('AICore');
       const contextRows = records.filter((r) => !response.some((rr) => matchRow(r, rr)));
       for (const row of response) {
diff --git a/srv/AICoreService.cds b/srv/AICoreService.cds
index 143733a..4c85bf9 100644
--- a/srv/AICoreService.cds
+++ b/srv/AICoreService.cds
@@ -226,7 +226,8 @@ service AICore {
       name                   : String;
       prediction_placeholder : String;
       task_type              : String enum {
-        classification = 'classification'
+        classification = 'classification';
+        regression     = 'regression';
       }
     }
   },
diff --git a/srv/AICoreService.js b/srv/AICoreService.js
index 482314c..c995967 100644
--- a/srv/AICoreService.js
+++ b/srv/AICoreService.js
@@ -20,6 +20,13 @@ import { createConfiguration, readConfigurations } from './ai-core/configuration
 
 const LOG = cds.log('@cap-js/ai');
 
+// RPT-1 inference only accepts dtype values 'string', 'numeric' or 'date'.
+// - Booleans round-trip as 'true' / 'false' strings.
+// - cds.Date maps to 'date' (calendar date, no time portion).
+// - cds.DateTime / cds.Timestamp keep their time portion: declared as
+//   'string' so RPT-1 treats the ISO value as an opaque categorical token
+//   instead of date-parsing it (which would drop the time and may reject
+//   ISO timestamps that aren't pure YYYY-MM-DD).
 const CDS_TO_PYTHON_DTYPE = {
   'cds.String': 'string',
   'cds.LargeString': 'string',
@@ -32,17 +39,51 @@ const CDS_TO_PYTHON_DTYPE = {
   'cds.UInt8': 'numeric',
   'cds.Decimal': 'numeric',
   'cds.Double': 'numeric',
-  'cds.Boolean': 'bool',
+  'cds.Boolean': 'string',
   'cds.Date': 'date',
   'cds.Time': 'string',
-  'cds.DateTime': 'datetime',
-  'cds.Timestamp': 'datetime'
+  'cds.DateTime': 'string',
+  'cds.Timestamp': 'string'
 };
 
 function cdsToPythonDtype(cdsType) {
   return CDS_TO_PYTHON_DTYPE[cdsType];
 }
 
+const NUMERIC_CDS_TYPES = new Set([
+  'cds.Integer',
+  'cds.Integer64',
+  'cds.Int16',
+  'cds.Int32',
+  'cds.Int64',
+  'cds.UInt8',
+  'cds.Decimal',
+  'cds.Double'
+]);
+
+// Pick the RPT-1 task type per target column. Fields with a ValueList always
+// get `classification` — they represent categorical choices. Numeric scalars
+// opted in via `@UI.RecommendationState` (without a ValueList) get `regression`
+// so the model can interpolate continuous values; everything else gets
+// `classification`.
+function pickTaskType(entity, columnName) {
+  const ele = entity?.elements?.[columnName];
+  if (!ele) return 'classification';
+  const hasValueList =
+    ele['@Common.ValueList.CollectionPath'] ||
+    cds.model.definitions[ele.target]?.['@cds.odata.valuelist'];
+  if (hasValueList) return 'classification';
+  if (ele['@UI.RecommendationState'] && NUMERIC_CDS_TYPES.has(ele.type)) return 'regression';
+  return 'classification';
+}
+
+// RPT-1 inference limits, per
+// https://help.sap.com/docs/sap-ai-core/generative-ai/sap-rpt-1
+// Exceeding either causes HTTP 422; we warn and skip the prediction so the
+// surrounding READ still completes instead of breaking the whole response.
+const RPT1_MAX_TARGET_COLUMNS = 10;
+const RPT1_MAX_ROW_COLUMNS = 100;
+
 export default class AICore extends cds.ApplicationService {
   init() {
     this.on('fetchPredictions', this._fetchPrediction);
@@ -103,6 +144,25 @@ export default class AICore extends cds.ApplicationService {
 
   async _fetchPrediction(req) {
     const { rows, entity: entityName, predictionColumns } = req.data;
+    // Empty rows would crash the schema-derivation reduce (rows[0] is
+    // undefined). Happens routinely when a draft composition is being
+    // read with no active rows yet — there is nothing to predict from.
+    if (!rows?.length) return {};
+    if (predictionColumns.length > RPT1_MAX_TARGET_COLUMNS) {
+      LOG.warn(
+        `Skipping recommendations for ${entityName}: ${predictionColumns.length} target columns exceeds the RPT-1 limit of ${RPT1_MAX_TARGET_COLUMNS}. ` +
+          'Opt fields out via @UI.RecommendationState : 0 to bring the count down.'
+      );
+      return {};
+    }
+    const rowColumnCount = Object.keys(rows[0]).length;
+    if (rowColumnCount > RPT1_MAX_ROW_COLUMNS) {
+      LOG.warn(
+        `Skipping recommendations for ${entityName}: rows carry ${rowColumnCount} columns, exceeding the RPT-1 limit of ${RPT1_MAX_ROW_COLUMNS}. ` +
+          'Either narrow the entity projection or opt out @cds.api.ignore-style columns that are not useful as features.'
+      );
+      return {};
+    }
     const entity = (cds.context?.model ?? cds.model).definitions[entityName];
     const dataSchema = entity
       ? Object.keys(entity.elements).reduce((acc, ele) => {
@@ -123,7 +183,7 @@ export default class AICore extends cds.ApplicationService {
           target_columns: predictionColumns.map((c) => ({
             name: c,
             prediction_placeholder: '[PREDICT]',
-            task_type: 'classification'
+            task_type: pickTaskType(entity, c)
           }))
         },
         // SAP_RECOMMENDATIONS_ID is generated in case the entity has composed keys or a key not named ID