feat(n8n): add Ollama embeddings to duplicate detector

implementations/n8n/duplicate-issue-detector/README.md

@@ -21,12 +21,12 @@ This workflow ships as **two separate flows in one file**:
 **1. Backfill (manual trigger, run once)**
 ```
 Manual Trigger → Get All Open Issues → Loop (1 at a time)
-   → Embed (OpenAI) → Index into Qdrant → Wait 1s → next issue
+   → Embed (OpenAI or Ollama) → Index into Qdrant → Wait 1s → next issue
 ```
 
 **2. Live Detection (GitHub trigger, runs forever after)**
 ```
-New Issue Opened → filter action=opened → Embed (OpenAI)
+New Issue Opened → filter action=opened → Embed (OpenAI or Ollama)
    → Search Qdrant for closest match → score ≥ 0.87?
         → yes: comment with matched issue # and score
         → either way: index this issue into Qdrant
@@ -41,26 +41,41 @@ The backfill flow exists because duplicate detection is useless with an empty ve
 ### Prerequisites
 - **n8n instance** — self-hosted or Cloud
 - **GitHub account** — fine-grained Personal Access Token scoped to one repo, with **Issues: Read & Write** permission
-- **OpenAI API key** — used only for embeddings (`text-embedding-3-small`), which cost a fraction of a cent per issue — not the same cost tier as chat completion models
+- **Embedding provider** — OpenAI (`text-embedding-3-small`) by default, or a self-hosted Ollama server running `nomic-embed-text`
 - **Qdrant instance** — free to self-host via Docker, or use Qdrant Cloud's free tier
 
 ### Installation
 1. **Import the workflow** — n8n: **Workflows → Import from File**, select `workflow.json`
-2. **Set credentials** — GitHub PAT, OpenAI API key
+2. **Set credentials** — GitHub PAT and, when using the default provider, an OpenAI API key
 3. **Fill in the placeholders** (table below)
 
+### Local Ollama option
+
+Both flows include a disabled Ollama HTTP Request node alongside the enabled OpenAI node. To use local embeddings:
+
+1. Install [Ollama](https://ollama.com), then pull the default model:
+   ```bash
+   ollama pull nomic-embed-text
+   ```
+2. Set `REPLACE_OLLAMA_BASE_URL` in both Ollama nodes to the URL reachable from n8n (for example, `http://localhost:11434`). n8n Cloud cannot reach a service on your computer's `localhost`; use a self-hosted n8n instance or an appropriately exposed Ollama endpoint.
+3. In both the live and backfill paths, disable the OpenAI node and enable the corresponding Ollama node. Do not leave both providers enabled, or each issue will be processed twice.
+4. Recreate the Qdrant collection before switching providers. `nomic-embed-text` returns 768-dimensional vectors, so the collection must use size `768`; do not mix them with the existing 1536-dimensional OpenAI vectors. Run the backfill flow again after recreating the collection.
+
+To use another Ollama embedding model, change the model name in both Ollama nodes and create the Qdrant collection with that model's actual output dimension.
+
 ---
 
 ## ⚠️ Before You Trust It Running Unattended
 
 Skipping any of these three steps means the workflow will either fail outright or silently do nothing useful — do them in order, before turning on the live trigger:
 
-1. **Create the Qdrant collection first.** It needs to exist with the correct vector size before either flow will work — `text-embedding-3-small` produces 1536-dimensional vectors:
+1. **Create the Qdrant collection first.** It needs to exist with the correct vector size before either flow will work. The default OpenAI path uses 1536-dimensional `text-embedding-3-small` vectors:
    ```bash
    curl -X PUT "{qdrant_url}/collections/{collection_name}" \
      -H "Content-Type: application/json" \
      -d '{"vectors": {"size": 1536, "distance": "Cosine"}}'
    ```
+   If you switch to Ollama, use the local model's output dimension instead (768 for `nomic-embed-text`; see the local setup above).
 
 2. **Run the backfill flow once, manually, before enabling the live trigger.** This is the step most duplicate-detector tutorials skip entirely — without it, the live trigger has nothing to compare new issues against, and will run for weeks appearing to work while actually catching nothing.
 
@@ -75,6 +90,7 @@ Skipping any of these three steps means the workflow will either fail outright o
 | `REPLACE_OWNER` / `REPLACE_REPO` | 3 GitHub nodes | Your repo owner/name |
 | GitHub credential | 3 GitHub nodes | Fine-grained PAT, Issues read+write only |
 | OpenAI credential | 2 embedding nodes | Your OpenAI API key |
+| `REPLACE_OLLAMA_BASE_URL` | 2 disabled Ollama nodes | Ollama base URL reachable from n8n, without the `/api` suffix |
 | `REPLACE_QDRANT_URL` | 4 HTTP Request nodes | Your Qdrant instance URL |
 | `REPLACE_COLLECTION_NAME` | Same 4 nodes | Any name, e.g. `github-issues` |
 | `REPLACE_QDRANT_API_KEY` | Same 4 nodes | Your Qdrant API key (Qdrant Cloud) or leave as-is for local Qdrant |
@@ -93,7 +109,7 @@ Skipping any of these three steps means the workflow will either fail outright o
 
 - [ ] Auto-label flagged issues (e.g. `possible-duplicate`) in addition to commenting
 - [ ] Support closed issues in the similarity search, not just open ones
-- [ ] Local embedding option (Ollama `nomic-embed-text`) as a free, private alternative to OpenAI — documented as a swap-in, same pattern as this collection's other workflows
+- [x] Local embedding option (Ollama `nomic-embed-text`) as a free, private alternative to OpenAI — documented as a swap-in, same pattern as this collection's other workflows
 - [ ] Auto re-tune threshold suggestion based on maintainer feedback (thumbs up/down on the comment)
 
 ---

implementations/n8n/duplicate-issue-detector/workflow.json

@@ -73,7 +73,23 @@
     },
     {
       "parameters": {
-        "jsCode": "const vector = $input.first().json.data[0].embedding;\nconst prev = $('Prepare Issue Text').first().json;\nreturn [{ json: { ...prev, vector } }];"
+        "method": "POST",
+        "url": "=REPLACE_OLLAMA_BASE_URL/api/embed",
+        "sendBody": true,
+        "specifyBody": "json",
+        "jsonBody": "={{ JSON.stringify({ model: \"nomic-embed-text\", input: $json.combinedText }) }}",
+        "options": {}
+      },
+      "id": "embed-ollama-live",
+      "name": "Embed New Issue (Ollama)",
+      "type": "n8n-nodes-base.httpRequest",
+      "typeVersion": 4.2,
+      "position": [-1040, -80],
+      "disabled": true
+    },
+    {
+      "parameters": {
+        "jsCode": "const response = $input.first().json;\nconst vector = response.data?.[0]?.embedding || response.embeddings?.[0];\nif (!Array.isArray(vector)) throw new Error('Embedding response did not contain a vector');\nconst prev = $('Prepare Issue Text').first().json;\nreturn [{ json: { ...prev, vector } }];"
       },
       "id": "extract-vector-live",
       "name": "Extract Vector",
@@ -237,7 +253,23 @@
     },
     {
       "parameters": {
-        "jsCode": "const vector = $input.first().json.data[0].embedding;\nconst prev = $('Prepare Issue Text (Backfill)').first().json;\nreturn [{ json: { ...prev, vector } }];"
+        "method": "POST",
+        "url": "=REPLACE_OLLAMA_BASE_URL/api/embed",
+        "sendBody": true,
+        "specifyBody": "json",
+        "jsonBody": "={{ JSON.stringify({ model: \"nomic-embed-text\", input: $json.combinedText }) }}",
+        "options": {}
+      },
+      "id": "embed-ollama-backfill",
+      "name": "Embed Issue (Ollama)",
+      "type": "n8n-nodes-base.httpRequest",
+      "typeVersion": 4.2,
+      "position": [-820, 520],
+      "disabled": true
+    },
+    {
+      "parameters": {
+        "jsCode": "const response = $input.first().json;\nconst vector = response.data?.[0]?.embedding || response.embeddings?.[0];\nif (!Array.isArray(vector)) throw new Error('Embedding response did not contain a vector');\nconst prev = $('Prepare Issue Text (Backfill)').first().json;\nreturn [{ json: { ...prev, vector } }];"
       },
       "id": "extract-vector-backfill",
       "name": "Extract Vector (Backfill)",
@@ -281,11 +313,17 @@
       "main": [[{ "node": "Prepare Issue Text", "type": "main", "index": 0 }]]
     },
     "Prepare Issue Text": {
-      "main": [[{ "node": "Embed New Issue (OpenAI)", "type": "main", "index": 0 }]]
+      "main": [[
+        { "node": "Embed New Issue (OpenAI)", "type": "main", "index": 0 },
+        { "node": "Embed New Issue (Ollama)", "type": "main", "index": 0 }
+      ]]
     },
     "Embed New Issue (OpenAI)": {
       "main": [[{ "node": "Extract Vector", "type": "main", "index": 0 }]]
     },
+    "Embed New Issue (Ollama)": {
+      "main": [[{ "node": "Extract Vector", "type": "main", "index": 0 }]]
+    },
     "Extract Vector": {
       "main": [[{ "node": "Search Similar Issues (Qdrant)", "type": "main", "index": 0 }]]
     },
@@ -317,11 +355,17 @@
       ]
     },
     "Prepare Issue Text (Backfill)": {
-      "main": [[{ "node": "Embed Issue (OpenAI)", "type": "main", "index": 0 }]]
+      "main": [[
+        { "node": "Embed Issue (OpenAI)", "type": "main", "index": 0 },
+        { "node": "Embed Issue (Ollama)", "type": "main", "index": 0 }
+      ]]
     },
     "Embed Issue (OpenAI)": {
       "main": [[{ "node": "Extract Vector (Backfill)", "type": "main", "index": 0 }]]
     },
+    "Embed Issue (Ollama)": {
+      "main": [[{ "node": "Extract Vector (Backfill)", "type": "main", "index": 0 }]]
+    },
     "Extract Vector (Backfill)": {
       "main": [[{ "node": "Index Issue (Qdrant Upsert)", "type": "main", "index": 0 }]]
     },

tests/test_duplicate_issue_detector_workflow.py

@@ -0,0 +1,56 @@
+import json
+from pathlib import Path
+
+ROOT = Path(__file__).resolve().parents[1]
+WORKFLOW_PATH = ROOT / "implementations" / "n8n" / "duplicate-issue-detector" / "workflow.json"
+README_PATH = ROOT / "implementations" / "n8n" / "duplicate-issue-detector" / "README.md"
+
+
+def _workflow() -> dict:
+    return json.loads(WORKFLOW_PATH.read_text(encoding="utf-8"))
+
+
+def _node(workflow: dict, name: str) -> dict:
+    return next(node for node in workflow["nodes"] if node["name"] == name)
+
+
+def _targets(workflow: dict, source: str) -> set[str]:
+    return {
+        edge["node"]
+        for output in workflow["connections"][source]["main"]
+        for edge in output
+    }
+
+
+def test_duplicate_detector_supports_swappable_ollama_embeddings():
+    workflow = _workflow()
+    live_ollama = _node(workflow, "Embed New Issue (Ollama)")
+    backfill_ollama = _node(workflow, "Embed Issue (Ollama)")
+
+    for node in (live_ollama, backfill_ollama):
+        assert node["type"] == "n8n-nodes-base.httpRequest"
+        assert node["disabled"] is True
+        assert node["parameters"]["url"] == "=REPLACE_OLLAMA_BASE_URL/api/embed"
+        assert 'model: "nomic-embed-text"' in node["parameters"]["jsonBody"]
+
+    assert _targets(workflow, "Prepare Issue Text") == {
+        "Embed New Issue (OpenAI)",
+        "Embed New Issue (Ollama)",
+    }
+    assert _targets(workflow, "Prepare Issue Text (Backfill)") == {
+        "Embed Issue (OpenAI)",
+        "Embed Issue (Ollama)",
+    }
+    assert _targets(workflow, "Embed New Issue (Ollama)") == {"Extract Vector"}
+    assert _targets(workflow, "Embed Issue (Ollama)") == {"Extract Vector (Backfill)"}
+
+    for name in ("Extract Vector", "Extract Vector (Backfill)"):
+        code = _node(workflow, name)["parameters"]["jsCode"]
+        assert "data?.[0]?.embedding" in code
+        assert "embeddings?.[0]" in code
+        assert "Embedding response did not contain a vector" in code
+
+    readme = README_PATH.read_text(encoding="utf-8")
+    assert "REPLACE_OLLAMA_BASE_URL" in readme
+    assert "768" in readme
+    assert "Do not leave both providers enabled" in readme