CogStack · jocelyneholdbrook · Apr 27, 2026
diff --git a/cogstack-cohorter/NL2DSL/Dockerfile b/cogstack-cohorter/NL2DSL/Dockerfile
@@ -3,7 +3,7 @@ FROM node:20-alpine
 ENV NODE_ENV=production \
     PORT=3002 \
     OLLAMA_URL=http://ollama:11434/api/generate \
-    OLLAMA_MODEL=gpt-oss:20b \
+    OLLAMA_MODEL=llama3.2:3b \
     MEDCAT_URL=http://cohorter-medcat:5000 \
     ALLOW_ORIGINS=*
 

diff --git a/cogstack-cohorter/NL2DSL/README.md b/cogstack-cohorter/NL2DSL/README.md
@@ -45,7 +45,7 @@ Run:
 docker run -d --name cohorter-nl2dsl \
   -p 3002:3002 \
   -e OLLAMA_URL="http://ollama:11434/api/generate" \
-  -e OLLAMA_MODEL="gpt-oss:20b" \
+  -e OLLAMA_MODEL="llama3.2:3b" \
   -e MEDCAT_URL="http://cohorter-medcat:3001" \
   -e ALLOW_ORIGINS="*" \
   --restart unless-stopped \
@@ -64,7 +64,7 @@ NL2DSL reads the following environment variables (as used in `docker_run.sh` / y
   Ollama generate endpoint URL, e.g. `http://ollama:11434/api/generate`
 
 - `OLLAMA_MODEL`  
-  Model name to use in Ollama, e.g. `gpt-oss:20b`
+  Model name to use in Ollama, e.g. `llama3.2:3b`
 
 - `MEDCAT_URL`  
   MedCAT API base URL, e.g. `http://cohorter-medcat:5000`
@@ -121,7 +121,7 @@ node server.js
 Pull the model inside the ollama container (example):
 
 ```bash
-docker exec -it ollama ollama pull gpt-oss:20b
+docker exec -it ollama ollama pull llama3.2:3b
 ```
 
 (Replace with your actual model name.)
diff --git a/cogstack-cohorter/NL2DSL/docker_run.sh b/cogstack-cohorter/NL2DSL/docker_run.sh
@@ -6,7 +6,7 @@ sudo docker network connect cohorter-net cohorter-medcat
 sudo docker run -d --name cohorter-nl2dsl --network cohorter-net \
   -p 3002:3002 \
   -e OLLAMA_URL="http://ollama:11434/api/generate" \
-  -e OLLAMA_MODEL="gpt-oss:20b" \
+  -e OLLAMA_MODEL="llama3.2:3b" \
   -e MEDCAT_URL="http://cohorter-medcat:5000" \
   -e ALLOW_ORIGINS="*" \
   --restart unless-stopped \

diff --git a/cogstack-cohorter/NL2DSL/server.js b/cogstack-cohorter/NL2DSL/server.js
@@ -7,7 +7,7 @@ import cors from 'cors';
 // ---- config ----
 const PORT = process.env.PORT || 3002;
 const OLLAMA_URL = process.env.OLLAMA_URL || 'http://127.0.0.1:8002/api/generate';
-const MODEL = process.env.OLLAMA_MODEL || 'gpt-oss:20b'; // 'phi4:latest';
+const MODEL = process.env.OLLAMA_MODEL || 'llama3.2:3b';
 
 // CORS: allow list via env (comma-separated) or "*"
 const allowList = (process.env.ALLOW_ORIGINS || '*')

diff --git a/cogstack-cohorter/README.md b/cogstack-cohorter/README.md
@@ -53,7 +53,7 @@ docker compose up --build -d
 
 NL2DSL uses:
 - `OLLAMA_URL` (default in compose: `http://ollama:11434/api/generate`)
-- `OLLAMA_MODEL` (default: `gpt-oss:20b`)
+- `OLLAMA_MODEL` (default: `llama3.2:3b`)
 - `MEDCAT_URL` (default: `http://cohorter-medcat:5000`)
 - `ALLOW_ORIGINS` (default: `*`)
 

diff --git a/cogstack-cohorter/WebAPP/client-react/public/assets/js/init-alpine.js b/cogstack-cohorter/WebAPP/client-react/public/assets/js/init-alpine.js
@@ -1,7 +1,6 @@
 function data() {
 
-  const NL2DSL_URL   = '/nl2dsl';
-  const NL2DSL_MODEL = 'gpt-oss:20b';
+  const NL2DSL_URL = '/nl2dsl';
 
   // read version from URL
   const url = new URL(window.location.href);
@@ -346,13 +345,12 @@ function data() {
       this.nl_loading = true;
 
       try {
-        // 1) NL -> DSL (force gpt-oss-20b)
+        // 1) NL -> DSL — model is configured server-side via OLLAMA_MODEL env var
         const compileResp = await fetch(NL2DSL_URL, {
           method: 'POST',
           headers: { 'Content-Type': 'application/json' },
           body: JSON.stringify({
-            query: q,
-            model: NL2DSL_MODEL         // ← use gpt-oss-20b
+            query: q
             // If your API also accepts temperature/top_p etc., add them here.
             // e.g. temperature: 0.2
           })

diff --git a/cogstack-cohorter/WebAPP/client-react/src/alpine-data.js b/cogstack-cohorter/WebAPP/client-react/src/alpine-data.js
@@ -1,7 +1,6 @@
 export function createAppState() {
 
-  const NL2DSL_URL   = '/nl2dsl';
-  const NL2DSL_MODEL = 'gpt-oss:20b';
+  const NL2DSL_URL = '/nl2dsl';
   const deepClone = (value) => {
     if (value == null) return value;
     if (typeof structuredClone === 'function') {
@@ -358,13 +357,12 @@ export function createAppState() {
       this.nl_loading = true;
 
       try {
-        // 1) NL -> DSL (force gpt-oss-20b)
+        // 1) NL -> DSL — model is configured server-side via OLLAMA_MODEL env var
         const compileResp = await fetch(NL2DSL_URL, {
           method: 'POST',
           headers: { 'Content-Type': 'application/json' },
           body: JSON.stringify({
-            query: q,
-            model: NL2DSL_MODEL         // ← use gpt-oss-20b
+            query: q
             // If your API also accepts temperature/top_p etc., add them here.
             // e.g. temperature: 0.2
           })

diff --git a/cogstack-cohorter/WebAPP/client/public/assets/js/init-alpine.js b/cogstack-cohorter/WebAPP/client/public/assets/js/init-alpine.js
@@ -1,7 +1,6 @@
 function data() {
 
-  const NL2DSL_URL   = '/nl2dsl';
-  const NL2DSL_MODEL = 'gpt-oss:20b';
+  const NL2DSL_URL = '/nl2dsl';
 
   // read version from URL
   const url = new URL(window.location.href);
@@ -346,13 +345,12 @@ function data() {
       this.nl_loading = true;
 
       try {
-        // 1) NL -> DSL (force gpt-oss-20b)
+        // 1) NL -> DSL — model is configured server-side via OLLAMA_MODEL env var
         const compileResp = await fetch(NL2DSL_URL, {
           method: 'POST',
           headers: { 'Content-Type': 'application/json' },
           body: JSON.stringify({
-            query: q,
-            model: NL2DSL_MODEL         // ← use gpt-oss-20b
+            query: q
             // If your API also accepts temperature/top_p etc., add them here.
             // e.g. temperature: 0.2
           })

diff --git a/cogstack-cohorter/WebAPP/server/server.js b/cogstack-cohorter/WebAPP/server/server.js
@@ -716,7 +716,7 @@ app.post('/nl2dsl', async (req, res) => {
         const r = await fetch(NL2DSL_SERVER, {
             method: 'POST',
             headers: { 'Content-Type': 'application/json' },
-            body: JSON.stringify(req.body),   // e.g. { query: "...", model: "gpt-oss-20b" }
+            body: JSON.stringify(req.body),   // e.g. { query: "...", model: "llama3.2:3b" }
         });
         const text = await r.text();
         console.log(text);

diff --git a/cogstack-cohorter/docker-compose.yml b/cogstack-cohorter/docker-compose.yml
@@ -16,7 +16,7 @@ services:
         until ollama list > /dev/null 2>&1; do
           echo '[ollama-pull] Waiting for Ollama...'; sleep 2;
         done &&
-        ollama pull ${OLLAMA_MODEL:-gpt-oss:20b}
+        ollama pull ${OLLAMA_MODEL:-llama3.2:3b}
       "
     environment:
       OLLAMA_HOST: "http://ollama:11434"
@@ -44,7 +44,7 @@ services:
       - "3002:3002"
     environment:
       OLLAMA_URL: "http://ollama:11434/api/generate"
-      OLLAMA_MODEL: "gpt-oss:20b"
+      OLLAMA_MODEL: "llama3.2:3b"
       MEDCAT_URL: "http://cohorter-medcat:5000"
       ALLOW_ORIGINS: "*"
     depends_on:

diff --git a/helm-charts/cogstack-cohorter-helm/README.md b/helm-charts/cogstack-cohorter-helm/README.md
@@ -19,7 +19,7 @@ MedCAT and Ollama are deployed as **subcharts**:
 
 - Kubernetes 1.21+
 - Helm 3.10+
-- Sufficient node resources for the Ollama model (the default `gpt-oss:20b` requires ~14 GB of memory/VRAM)
+- Sufficient node resources for the Ollama model. The default `llama3.2:3b` (~2 GB) runs on CPU. For better NL → DSL accuracy, consider a larger model — see [Ollama model library](https://ollama.com/library)
 
 ## Installation
 
@@ -41,14 +41,16 @@ ollama:
   ollama:
     models:
       pull:
-        - gpt-oss:20b   # pulled automatically on first startup
+        - llama3.2:3b   # pulled automatically on first startup
   persistentVolume:
     enabled: true
     size: 10Gi
 ```
 
 Models are pulled automatically by the otwld subchart's built-in init container. Change `ollama.ollama.models.pull` to use a different model — make sure `nl2dsl.env.OLLAMA_MODEL` matches.
 
+`llama3.2:3b` is a lightweight default that runs on CPU but produces lower-quality NL → DSL results. For better accuracy, upgrade to a larger model such as `llama3.1:8b`, `llama3.3:70b`, or `gpt-oss:20b`. Browse all available models at [https://ollama.com/library](https://ollama.com/library).
+
 ### MedCAT
 
 ```yaml
@@ -161,7 +163,7 @@ For issues and questions, please visit the [CogStack GitHub repository](https://
 | nl2dsl.affinity | object | `{}` |  |
 | nl2dsl.enabled | bool | `true` |  |
 | nl2dsl.env.ALLOW_ORIGINS | string | `"*"` |  |
-| nl2dsl.env.OLLAMA_MODEL | string | `"gpt-oss:20b"` |  |
+| nl2dsl.env.OLLAMA_MODEL | string | `"llama3.2:3b"` |  |
 | nl2dsl.image.pullPolicy | string | `"IfNotPresent"` |  |
 | nl2dsl.image.repository | string | `"cogstacksystems/cogstack-cohorter-nl2dsl"` |  |
 | nl2dsl.image.tag | string | `"latest"` |  |
@@ -180,7 +182,7 @@ For issues and questions, please visit the [CogStack GitHub repository](https://
 | nl2dsl.service.type | string | `"ClusterIP"` |  |
 | nl2dsl.tolerations | list | `[]` |  |
 | ollama.enabled | bool | `true` |  |
-| ollama.ollama.models.pull[0] | string | `"gpt-oss:20b"` |  |
+| ollama.ollama.models.pull[0] | string | `"llama3.2:3b"` |  |
 | ollama.persistentVolume.enabled | bool | `true` |  |
 | ollama.persistentVolume.size | string | `"10Gi"` |  |
 | ollama.persistentVolume.storageClass | string | `""` |  |

diff --git a/helm-charts/cogstack-cohorter-helm/values.yaml b/helm-charts/cogstack-cohorter-helm/values.yaml
@@ -188,7 +188,11 @@ nl2dsl:
   # OLLAMA_URL and MEDCAT_URL are set automatically from subchart service names.
   env:
     # Ollama model to use for NL → DSL compilation.
-    OLLAMA_MODEL: "gpt-oss:20b"
+    # llama3.2:3b is a lightweight default that runs on CPU but produces lower-quality results.
+    # For better NL → DSL accuracy, upgrade to a larger model — e.g. llama3.1:8b, llama3.3:70b,
+    # or gpt-oss:20b. Must match ollama.ollama.models.pull below.
+    # Browse available models at: https://ollama.com/library
+    OLLAMA_MODEL: "llama3.2:3b"
     # CORS origins allowed to call the NL2DSL API.
     ALLOW_ORIGINS: "*"
 
@@ -263,9 +267,11 @@ ollama:
   ollama:
     # Models pulled automatically by the built-in init container on first startup.
     # Ensure nl2dsl.env.OLLAMA_MODEL matches the model listed here.
+    # For better results, replace llama3.2:3b with a larger model — e.g. llama3.1:8b,
+    # llama3.3:70b, or gpt-oss:20b. Browse options at: https://ollama.com/library
     models:
       pull:
-        - gpt-oss:20b
+        - llama3.2:3b
 
   persistentVolume:
     enabled: true