Ollama

.env

@@ -4,4 +4,6 @@
 # Prisma supports the native connection string format for PostgreSQL, MySQL, SQLite, SQL Server, MongoDB and CockroachDB.
 # See the documentation for all the connection string options: https://pris.ly/d/connection-strings
 
-DATABASE_URL="postgresql://johndoe:randompassword@localhost:5432/mydb?schema=public"
+DATABASE_URL="postgresql://watches:watches@localhost:${FORWARD_DB_PORT:-5432}/watches?schema=public"
+EMBEDDING_MODEL=all-minilm
+CHAT_MODEL=phi3:mini

.upsun/config.yaml

@@ -6,19 +6,46 @@ services:
         - vector
 
 applications:
+  ollama:
+    stack:
+      - "python@3.12"
+      - "ollama"
+    container_profile: BALANCED
+    relationships:
+      db: "postgresql:postgresql"
+    mounts:
+      "/.ollama":
+        source: "storage"
+        source_path: ".ollama"
+    web:
+      commands:
+        start: "OLLAMA_HOST=:$PORT ollama serve"
+    hooks:
+      build: |
+        set -eux
+        cd vectorize/
+        pip install -r requirements.txt
+      deploy: |
+        set -eux
+        set -a
+        . /app/.env
+        export OLLAMA_HOST=:$PORT
+        ollama pull $EMBEDDING_MODEL
+        ollama pull $CHAT_MODEL
+
+        cd vectorize/
+        python vectorize.py
+
   app:
     stack:
       - "nodejs@22"
-      - "python@3.12"
     relationships:
       db: "postgresql:postgresql"
+      ollama: "ollama:http"
     mounts:
       "/.npm":
         source: "storage"
         source_path: "npm"
-      "/.cache":
-        source: "storage"
-        source_path: "cache"
     web:
       commands:
         start: "npx next start -p $PORT"
@@ -28,12 +55,6 @@ applications:
         npm install
         npx prisma generate
         npm run build
-        cd vectorize/
-        pip install -r requirements.txt
-      post_deploy: |
-        set -eux
-        cd vectorize/
-        python vectorize.py
 
 routes:
   "https://{default}/":

README.md

@@ -6,8 +6,8 @@
 
 The data located in `vectorize/data` is a [dataset from Kaggle](https://www.kaggle.com/datasets/rkiattisak/luxury-watches-price-dataset/data).
 
-This example uses 3 models from HuggingFace:
-- `vectorize.py` uses the `sentence-transformers` and run `feature_extractor` with the `all-MiniLM-L6-v2` model locally (CPU). The model is downloaded (88MB) on the first run in the `.cache` folder. The embeddings are then stored in the postgres database. `vectorize.py` also migrates the database schema if it does not exists.
+This example uses a local Ollama server:
+- `vectorize.py` uses the `all-minilm` model run `feature_extractor` with the `all-MiniLM-L6-v2` model locally (CPU). The model is downloaded (88MB) on the first run in the `~/.cache/huggingface/` folder. The embeddings are then stored in the postgres database. `vectorize.py` also migrates the database schema if it does not exist.
 - When the user inputs a prompt, the `recommend` action will first run `featureExtraction` on the serverless HuggingFace inference API with the same model `all-MiniLM-L6-v2` to get the embedding.
 - The `recommend` action then query the database using the vector to get 5 similar watches.
 - The last step of the action is to query a text generation model on HuggingFace inference API to generate a proper text answer. We are using `mistralai/Mistral-7B-Instruct-v0.2` in this example.
@@ -22,15 +22,17 @@ The `openai` branch does the same but relies on the OpenAI API for all the LLM a
 cd vectorize
 python3 -m venv venv
 source ./venv/bin/activate
-pip install -r requirements.txt (or requirements-mac.txt)
+pip install -r requirements.txt
 ```
 
 ### Install dependencies and run services
 
 ```
-docker-compose up
+export FORWARD_DB_PORT=8090
+docker compose up -d
 cd vectorize/
-DB_PATH=watches DB_HOST=127.0.0.1 DB_USERNAME=watches DB_PASSWORD=watches HUGGINGFACE_TOKEN=hf_****** python3 vectorize.py
+source ../.env.local
+DB_PORT=$FORWARD_DB_PORT DB_PATH=watches DB_HOST=127.0.0.1 DB_USERNAME=watches DB_PASSWORD=watches python3 vectorize.py
 cd ../
 npm install
 npm rum dev
@@ -40,8 +42,9 @@ npm rum dev
 
 ```
 upsun project:create
-upsun variable:create --name HUGGINGFACE_TOKEN --prefix env: --level project
-upsun push
+upsun variable:create --name env:EMBEDDING_MODEL --level project --visible-build=true --value=all-minilm
+upsun variable:create --name env:CHAT_MODEL --level project --visible-build=true --value=tinyllama:1.1b-chat-v1-q2_K
+upsun push --resources-init=manual
 ```
 
-The `vectorize.py` script is included in the `deploy` hook meaning that it will be triggered on every deploy. This is for demo purposes. You can run it manually instead to avoid delays in deployments.
+The `vectorize.py` script is included in the `deploy` hook meaning that it will be triggered on every deploy. This is for demo purposes. You can run it manually instead to avoid delays in deployments.

docker/pgsql/postgres.Dockerfile

@@ -1,4 +1,4 @@
-FROM postgres:latest
+FROM postgres:16
 
 RUN apt-get update && apt-get install -y \
     build-essential \