n8n-nodes-berget-mk 0.4.13 → 0.4.15

package/README.md

@@ -4,10 +4,10 @@ n8n community nodes for [Berget AI](https://berget.ai), packaged as a single ins
 
 Four nodes:
 
-- **Berget AI** — multi-resource action node for one-shot calls: **Chat** (completions, classification), **Image Analysis** (vision models), **Rerank** (document reranking), and **Speech to Text** (Swedish-tuned KB-Whisper). Can also be exposed as a tool to an AI Agent. (OCR is temporarily hidden — see [CHANGELOG.md](CHANGELOG.md) for `0.4.4` for details.)
+- **Berget AI** — multi-resource action node for one-shot calls. Resources: **Chat** (completions, classification, JSON Schema structured output), **Image Analysis** (vision-capable models), **Rerank** (document reranking), and **Speech to Text** (Swedish-tuned KB-Whisper, with optional diarization and word-level alignment). Can also be exposed as a tool to an AI Agent. (OCR is temporarily hidden — see [CHANGELOG.md](CHANGELOG.md) for `0.4.4` for details.)
 - **Berget AI Chat Model** — sub-node that plugs into n8n's built-in **AI Agent**, **Basic LLM Chain**, and other LangChain-based nodes. Exposes `reasoning_effort` and the full standard LLM parameter set.
 - **Berget AI Embeddings Model** — sub-node that plugs into n8n's **Vector Store** nodes (Supabase, Qdrant, Pinecone, PGVector, etc.) and **Question and Answer Chain**.
-- **Berget AI Reranker** — sub-node that plugs into Vector Store retrievers to reorder candidate documents by relevance before handing them to the chain or agent.
+- **Berget AI Reranker** — sub-node that plugs into Vector Store retrievers via the `AiReranker` connection, reordering candidates by relevance before they reach the agent or chain.
 
 > ⚠️ **Experimental — actively developed.** This package is pre-1.0 and may break between minor releases. Pin a specific version in production workflows until `1.0.0`. See [CHANGELOG.md](CHANGELOG.md) for breaking changes.
 
@@ -27,21 +27,35 @@ Then add a **Berget AI API** credential with your API key from [berget.ai](https
 
 1. Drop **Berget AI** onto the canvas, pick Resource = **Chat**, select a model, add a user message. Execute.
 
+For classification or structured extraction tasks, set **Options → Response Format = JSON Schema** and provide a schema. The model is forced to return parseable JSON conforming to your shape — no regex scraping of free-form text.
+
 ### Agent with tools and memory
 
 1. Add n8n's built-in **AI Agent**.
 2. Add **Berget AI Chat Model** and connect it to the Agent's Chat Model socket.
 3. Add Memory and Tool sub-nodes as needed — they work with Berget as the underlying LLM.
 
-### RAG / vector search
+### RAG with retrieval and reranking
 
-1. Add a Vector Store node (Supabase, Qdrant, etc.) or a Question and Answer Chain.
+1. Add a Vector Store node (Qdrant, Supabase, etc.) or a Question and Answer Chain.
 2. Add **Berget AI Embeddings Model** and connect it to the Embedding socket.
-3. Index documents or query as usual.
+3. Add **Berget AI Reranker** and connect it to the EmbeddingReranker socket — Vector Store will then retrieve a wider candidate set, the reranker reorders them by relevance, and only the best survive into the answer.
+4. Index documents or query as usual.
+
+### Image analysis
+
+1. Drop **Berget AI** onto the canvas, pick Resource = **Image Analysis**.
+2. Pick a vision-capable model (the dropdown is filtered automatically).
+3. Choose Input Type = **Binary File** (default — works with Form Trigger uploads, HTTP Request responses, etc.) or **Image URL**, and provide a **Text Input** prompt like `"Describe what you see"`.
+4. Execute.
 
-### Swedish speech transcription
+### Swedish speech transcription with speakers
 
-1. Drop **Berget AI** onto the canvas, pick Resource = **Speech to Text**, pick a model (defaults to `KB-Whisper-Large`), and point at an audio file.
+1. Drop **Berget AI** onto the canvas, pick Resource = **Speech to Text**.
+2. Provide the binary input data from a Form Trigger or HTTP Request.
+3. Optional: enable **Options → Diarize (Speaker Identification)** — the response will include a `speaker_transcript` field formatted as readable per-speaker paragraphs (`SPEAKER_00:\n...\n\nSPEAKER_01:\n...`), alongside the raw segment-level timestamps and word-level data.
+4. Optional: enable **Word-Level Alignment** for per-word timestamps useful in subtitle generation.
+5. Optional: add **Hotwords** (comma-separated) for proper nouns and domain vocabulary.
 
 ## Changelog
 

package/dist/nodes/BergetAi/chat.js

@@ -84,7 +84,7 @@ exports.chatProperties = [
                     { name: 'JSON Schema', value: 'json_schema' },
                 ],
                 default: 'text',
-                description: 'Force the model to return a specific response format. "JSON Object" tells the model to return any valid JSON. "JSON Schema" enforces a specific JSON schema you provide — set the schema with the JSON Schema fields below.',
+                description: 'Force the model to return a specific response format. "JSON Object" tells the model to return any valid JSON. "JSON Schema" enforces a specific JSON schema you provide — set the schema with the JSON Schema fields below. When either is selected, the parsed JSON is also exposed as a top-level "output" field on the node\'s output so downstream nodes (IF, Set, etc.) can reference its properties directly.',
             },
             {
                 displayName: 'JSON Schema Name',
@@ -131,6 +131,7 @@ exports.chatProperties = [
     },
 ];
 async function executeChat(context, itemIndex) {
+    var _a, _b;
     const credentials = await context.getCredentials('bergetAiApi');
     const model = context.getNodeParameter('chatModel', itemIndex);
     const messages = context.getNodeParameter('chatMessages.values', itemIndex, []);
@@ -173,5 +174,22 @@ async function executeChat(context, itemIndex) {
     if (status !== 200) {
         throw new n8n_workflow_1.NodeOperationError(context.getNode(), (0, shared_1.formatBergetError)('chat', status, data), { itemIndex });
     }
-    return data;
+    const result = data;
+    if (responseFormat === 'json_object' || responseFormat === 'json_schema') {
+        const choices = result.choices;
+        const rawContent = (_b = (_a = choices === null || choices === void 0 ? void 0 : choices[0]) === null || _a === void 0 ? void 0 : _a.message) === null || _b === void 0 ? void 0 : _b.content;
+        if (typeof rawContent === 'string' && rawContent.trim().length > 0) {
+            try {
+                const parsed = JSON.parse(rawContent);
+                if (parsed && typeof parsed === 'object') {
+                    result.output = parsed;
+                }
+            }
+            catch {
+                // Model returned non-JSON despite response_format being set.
+                // Leave output absent; raw string stays in choices[0].message.content.
+            }
+        }
+    }
+    return result;
 }

package/dist/nodes/BergetAi/speech.js

@@ -199,18 +199,42 @@ async function executeSpeech(context, itemIndex) {
     // segments and word timestamps. The raw segments/words/timestamps are
     // still preserved on the result object so power users can drill into
     // them when needed.
-    if (options.diarize &&
-        data &&
-        typeof data === 'object' &&
-        Array.isArray(data.segments)) {
-        const segments = data.segments;
-        const transcript = buildSpeakerTranscript(segments);
-        if (transcript) {
-            data.speaker_transcript = transcript;
+    if (options.diarize && data && typeof data === 'object') {
+        const segments = extractSegments(data);
+        if (segments) {
+            const transcript = buildSpeakerTranscript(segments);
+            if (transcript) {
+                data.speaker_transcript = transcript;
+            }
         }
     }
     return data;
 }
+/**
+ * Pull the segments array out of Berget's transcription response. Berget has
+ * been observed to return the segments in two different shapes:
+ *
+ *   Shape A (flat):   { segments: [...], language, text }
+ *   Shape B (nested): { segments: { segments: [...], ... }, language, text }
+ *
+ * Shape B is what the API returns for verbose_json with diarize=true (as of
+ * 2026-04). We check both so the speaker_transcript builder works regardless
+ * of which shape we get, and so future API changes that flatten or re-nest
+ * don't silently break the output.
+ */
+function extractSegments(data) {
+    const top = data.segments;
+    if (Array.isArray(top)) {
+        return top;
+    }
+    if (top && typeof top === 'object') {
+        const inner = top.segments;
+        if (Array.isArray(inner)) {
+            return inner;
+        }
+    }
+    return undefined;
+}
 /**
  * Build a "SPEAKER_00:\n...text...\n\nSPEAKER_01:\n..." style transcript
  * by walking the segments array, grouping consecutive segments that share

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "n8n-nodes-berget-mk",
-  "version": "0.4.13",
+  "version": "0.4.15",
   "description": "n8n community node for Berget AI. Multi-resource action node (chat, OCR, rerank, speech-to-text) plus Chat Model and Embeddings Model sub-nodes that plug into n8n's built-in AI Agent and Vector Store nodes.",
   "keywords": [
     "n8n-community-node-package",