graphile-llm 1.21.7 → 1.21.8

package/embed-inputs.js

@@ -134,13 +134,20 @@ function createEmbedIndex(graphql) {
         switch (rule.kind) {
             case 'nearby': {
                 delete next[fieldName];
-                if (isNonEmptyVector(value[rule.into]))
-                    return;
-                if (typeof text !== 'string') {
-                    throw new Error(`${typeName}: provide either \`${rule.into}\` or \`${fieldName}\`.`);
+                const hasVector = isNonEmptyVector(value[rule.into]);
+                const hasText = typeof text === 'string';
+                if (hasVector && hasText) {
+                    throw new Error(`${typeName}: provide exactly one of \`${rule.into}\` or \`${fieldName}\`, not both.`);
                 }
-                if (!embed)
+                if (!hasVector && !hasText) {
+                    throw new Error(`${typeName}: provide exactly one of \`${rule.into}\` or \`${fieldName}\`; neither was given.`);
+                }
+                if (hasVector)
                     return;
+                if (!embed) {
+                    throw new Error(`${typeName}: \`${fieldName}\` needs an embedding provider and none is configured; ` +
+                        `send \`${rule.into}\` instead.`);
+                }
                 const vector = await embed(text);
                 if (vector) {
                     next[rule.into] = vector;

package/esm/embed-inputs.js

@@ -126,13 +126,20 @@ export function createEmbedIndex(graphql) {
         switch (rule.kind) {
             case 'nearby': {
                 delete next[fieldName];
-                if (isNonEmptyVector(value[rule.into]))
-                    return;
-                if (typeof text !== 'string') {
-                    throw new Error(`${typeName}: provide either \`${rule.into}\` or \`${fieldName}\`.`);
+                const hasVector = isNonEmptyVector(value[rule.into]);
+                const hasText = typeof text === 'string';
+                if (hasVector && hasText) {
+                    throw new Error(`${typeName}: provide exactly one of \`${rule.into}\` or \`${fieldName}\`, not both.`);
                 }
-                if (!embed)
+                if (!hasVector && !hasText) {
+                    throw new Error(`${typeName}: provide exactly one of \`${rule.into}\` or \`${fieldName}\`; neither was given.`);
+                }
+                if (hasVector)
                     return;
+                if (!embed) {
+                    throw new Error(`${typeName}: \`${fieldName}\` needs an embedding provider and none is configured; ` +
+                        `send \`${rule.into}\` instead.`);
+                }
                 const vector = await embed(text);
                 if (vector) {
                     next[rule.into] = vector;

package/esm/plugins/text-search-plugin.d.ts

@@ -10,15 +10,15 @@
  * accepts a compound input (point + distance) and the plugin handles
  * the conversion to SQL internally.
  *
- * Clients provide `text` or `vector`; `vector` becomes optional. When both
- * are given the vector wins. `unifiedSearch` text is embedded too, so pgvector
+ * Clients provide exactly one of `text` or `vector`; `vector` becomes optional.
+ * Sending both, or neither, is an error. `unifiedSearch` text is embedded too, so pgvector
  * participates alongside the text adapters.
  *
  * The text is embedded while the operation is planned (see embed-inputs.ts),
  * before SQL is built.
  *
- * If no embedder is configured, `text` is ignored and `unifiedSearch` runs
- * on the text adapters only.
+ * If no embedder is configured, `text` on `VectorNearbyInput` is an error and
+ * `unifiedSearch` runs on the text adapters only.
  *
  * If the embedder returns null (e.g. quota exceeded when the metering
  * plugin is loaded), behavior depends on `onQuotaExceeded`:

package/esm/plugins/text-search-plugin.js

@@ -10,15 +10,15 @@
  * accepts a compound input (point + distance) and the plugin handles
  * the conversion to SQL internally.
  *
- * Clients provide `text` or `vector`; `vector` becomes optional. When both
- * are given the vector wins. `unifiedSearch` text is embedded too, so pgvector
+ * Clients provide exactly one of `text` or `vector`; `vector` becomes optional.
+ * Sending both, or neither, is an error. `unifiedSearch` text is embedded too, so pgvector
  * participates alongside the text adapters.
  *
  * The text is embedded while the operation is planned (see embed-inputs.ts),
  * before SQL is built.
  *
- * If no embedder is configured, `text` is ignored and `unifiedSearch` runs
- * on the text adapters only.
+ * If no embedder is configured, `text` on `VectorNearbyInput` is an error and
+ * `unifiedSearch` runs on the text adapters only.
  *
  * If the embedder returns null (e.g. quota exceeded when the metering
  * plugin is loaded), behavior depends on `onQuotaExceeded`:
@@ -138,7 +138,7 @@ export function createLlmTextSearchPlugin(options = {}) {
                         text: embedField({ kind: 'nearby', into: 'vector', onQuotaExceeded }, {
                             type: GraphQLString,
                             description: 'Natural language text to embed server-side for similarity search. ' +
-                                'Provide `text` or `vector`; if both are given, `vector` is used. ' +
+                                'Provide exactly one of `text` or `vector`; sending both or neither is an error. ' +
                                 'Requires the LLM plugin to be configured with an embedding provider.'
                         })
                     }, 'LlmTextSearchPlugin adding text field to VectorNearbyInput');
@@ -151,7 +151,7 @@ export function createLlmTextSearchPlugin(options = {}) {
                 },
                 finalize(schema, build) {
                     if (!build.llmEmbedder) {
-                        console.log('[graphile-llm] No embedder available — text on VectorNearbyInput is ignored ' +
+                        console.log('[graphile-llm] No embedder available — text on VectorNearbyInput is rejected ' +
                             'and unifiedSearch runs without pgvector. Configure an embedding provider to enable.');
                     }
                     return planTimeEmbedding(schema, build);

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "graphile-llm",
-  "version": "1.21.7",
+  "version": "1.21.8",
   "description": "LLM integration plugin for PostGraphile v5 — server-side text-to-vector embedding and text companion fields for pgvector columns",
   "author": "Constructive <developers@constructive.io>",
   "homepage": "https://github.com/constructive-io/constructive",
@@ -75,5 +75,5 @@
     "ollama",
     "openai"
   ],
-  "gitHead": "ca40d826b329e282aafba0f48f90cf35aac816e0"
+  "gitHead": "723c0a02471517c3015a218dd179fa9f53b47980"
 }

package/plugins/text-search-plugin.d.ts

@@ -10,15 +10,15 @@
  * accepts a compound input (point + distance) and the plugin handles
  * the conversion to SQL internally.
  *
- * Clients provide `text` or `vector`; `vector` becomes optional. When both
- * are given the vector wins. `unifiedSearch` text is embedded too, so pgvector
+ * Clients provide exactly one of `text` or `vector`; `vector` becomes optional.
+ * Sending both, or neither, is an error. `unifiedSearch` text is embedded too, so pgvector
  * participates alongside the text adapters.
  *
  * The text is embedded while the operation is planned (see embed-inputs.ts),
  * before SQL is built.
  *
- * If no embedder is configured, `text` is ignored and `unifiedSearch` runs
- * on the text adapters only.
+ * If no embedder is configured, `text` on `VectorNearbyInput` is an error and
+ * `unifiedSearch` runs on the text adapters only.
  *
  * If the embedder returns null (e.g. quota exceeded when the metering
  * plugin is loaded), behavior depends on `onQuotaExceeded`:

package/plugins/text-search-plugin.js

@@ -11,15 +11,15 @@
  * accepts a compound input (point + distance) and the plugin handles
  * the conversion to SQL internally.
  *
- * Clients provide `text` or `vector`; `vector` becomes optional. When both
- * are given the vector wins. `unifiedSearch` text is embedded too, so pgvector
+ * Clients provide exactly one of `text` or `vector`; `vector` becomes optional.
+ * Sending both, or neither, is an error. `unifiedSearch` text is embedded too, so pgvector
  * participates alongside the text adapters.
  *
  * The text is embedded while the operation is planned (see embed-inputs.ts),
  * before SQL is built.
  *
- * If no embedder is configured, `text` is ignored and `unifiedSearch` runs
- * on the text adapters only.
+ * If no embedder is configured, `text` on `VectorNearbyInput` is an error and
+ * `unifiedSearch` runs on the text adapters only.
  *
  * If the embedder returns null (e.g. quota exceeded when the metering
  * plugin is loaded), behavior depends on `onQuotaExceeded`:
@@ -142,7 +142,7 @@ function createLlmTextSearchPlugin(options = {}) {
                         text: (0, embed_inputs_1.embedField)({ kind: 'nearby', into: 'vector', onQuotaExceeded }, {
                             type: GraphQLString,
                             description: 'Natural language text to embed server-side for similarity search. ' +
-                                'Provide `text` or `vector`; if both are given, `vector` is used. ' +
+                                'Provide exactly one of `text` or `vector`; sending both or neither is an error. ' +
                                 'Requires the LLM plugin to be configured with an embedding provider.'
                         })
                     }, 'LlmTextSearchPlugin adding text field to VectorNearbyInput');
@@ -155,7 +155,7 @@ function createLlmTextSearchPlugin(options = {}) {
                 },
                 finalize(schema, build) {
                     if (!build.llmEmbedder) {
-                        console.log('[graphile-llm] No embedder available — text on VectorNearbyInput is ignored ' +
+                        console.log('[graphile-llm] No embedder available — text on VectorNearbyInput is rejected ' +
                             'and unifiedSearch runs without pgvector. Configure an embedding provider to enable.');
                     }
                     return (0, embed_inputs_1.planTimeEmbedding)(schema, build);