Update model references and dependencies to use Ollama embeddings

.github/ISSUE_TEMPLATE/bug_report.md

@@ -27,7 +27,7 @@ If applicable, add screenshots or log snippets.
 - **OS**: [e.g., Ubuntu 22.04, macOS 14]
 - **Python version**: [e.g., 3.13.1]
 - **Ollama version**: [e.g., 0.3.x]
-- **MedGemma model**: [e.g., MedAIBase/MedGemma1.5:4b]
+- **MedGemma model**: [e.g., medgemma:27b]
 - **Browser** (if UI issue): [e.g., Chrome 120]
 
 ## Additional Context

QUICKSTART.md

@@ -25,8 +25,9 @@ pip install -e ".[dev]"
 # 4. Set up pre-commit hooks
 pre-commit install
 
-# 5. Pull the MedGemma model (one-time)
-ollama pull MedAIBase/MedGemma1.5:4b
+# 5. Pull the models (one-time)
+ollama pull medgemma:27b
+ollama pull nomic-embed-text
 
 # 6. Run tests to verify setup
 python -m pytest tests/ -v

README.md

@@ -18,7 +18,7 @@ SELENE is a research prototype developed at HARMONI Lab, exploring privacy-prese
 - Daily Attune: capture rest/internal weather/clarity + notes; validated saves with backups.
 - Chat: contextualized queries, Chroma RAG, past-session recall, streaming MedGemma responses.
 - Clinical summary: deterministic stats/patterns/risk + single MedGemma call; PDF export via xhtml2pdf.
-- Local knowledge base: Chroma collections (medical_docs, chat_history) with SentenceTransformer embeddings.
+- Local knowledge base: Chroma collections (medical_docs, chat_history) with Ollama embeddings.
 - Safety: deterministic risk flags, conservative prompts, low temperature, offline defaults.
 
 ## Architecture (brief)
@@ -33,7 +33,7 @@ SELENE is a research prototype developed at HARMONI Lab, exploring privacy-prese
 
 ## Prerequisites
 - Python 3.11+
-- Ollama running locally with model `MedAIBase/MedGemma1.5:4b` pulled
+- Ollama running locally with model `medgemma:27b` pulled
 - Basic build deps for scientific stack (numpy/scipy/pandas) and xhtml2pdf; install via requirements.txt
 
 ## Quick Start
@@ -49,7 +49,7 @@ source med_env/bin/activate
 pip install -e ".[dev]"
 
 # 3. Pull model in Ollama (once)
-ollama pull MedAIBase/MedGemma1.5:4b
+ollama pull medgemma:27b
 
 # 4. Launch app
 streamlit run app.py
@@ -92,11 +92,11 @@ See [DIRECTORY_STRUCTURE.md](DIRECTORY_STRUCTURE.md) for the full tree.
 
 ## Configuration Highlights
 - Paths, model names, cache TTLs in [settings.py](src/selene/settings.py): RAG_TOP_K=2, contextualize cache 300s, RAG cache 600s, user context cache 180s.
-- Offline/telemetry disabled by default via envs set in settings (TRANSFORMERS_OFFLINE, HF_*_OFFLINE, CHROMA_TELEMETRY=False).
+- Offline/telemetry disabled by default via envs set in settings (CHROMA_TELEMETRY=False).
 - Logging defaults to DEBUG; file logging enabled by default (toggle LOG_TO_FILE).
 
 ## Knowledge Base Management
-- Chroma collections live under `data/user_data/user_med_db`; embeddings via SentenceTransformer all-MiniLM-L6-v2.
+- Chroma collections live under `data/user_data/user_med_db`; embeddings via Ollama nomic-embed-text.
 - Import/export and collection maintenance via [scripts/update_kb_chroma.py](scripts/update_kb_chroma.py) (keeps collection IDs stable).
 
 ## Safety & Guardrails

docs/technical_reference.md

@@ -111,7 +111,7 @@
 
 ### Vector-based retrieval (RAG)
 - Knowledge base stored in Chroma `medical_docs`.
-- Embeddings via `settings.get_embedding_function()` (SentenceTransformer).
+- Embeddings via `settings.get_embedding_function()` (OllamaEmbeddingFunction).
 - `med_logic.query_knowledge_base()` performs retrieval, formats chunks with `SOURCE` and `SECTION`, and returns (context, sources, full_results).
 - `update_kb_chroma.py` provides tools to import Chroma exports, clear collection contents safely, and keep collection IDs stable.
 
@@ -139,14 +139,14 @@
 
 ### What never leaves the device (by design)
 - User profile, pulse entries, chat history, and local knowledge base contents **are not transmitted externally** by the provided code paths.
-- Chroma telemetry is disabled (`CHROMA_TELEMETRY = False`). Transformers/HF offline flags are set in `settings.py`.
+- Chroma telemetry is disabled (`CHROMA_TELEMETRY = False`).
 
 ---
 
 ## 4. Reasoning Engine
 
 ### Model choice: MedGemma
-- Configured model: `settings.LLM_MODEL = "MedAIBase/MedGemma1.5:4b"`.
+- Configured model: `settings.LLM_MODEL = "medgemma:27b"`.
 - Rationale:
   - Domain-adapted medical reasoning (specialized training for menopause/clinical texts).
   - Size and local-run compatibility balance for edge devices when deployed via Ollama.
@@ -265,7 +265,7 @@
 
 ### Local-only processing guarantees
 - All user data (profile, pulse, chat) and the knowledge base (Chroma collections) are stored locally under `settings.USER_DATA_DIR` by default.
-- The LLM interaction is intended to occur to a local Ollama endpoint. `settings.py` sets `TRANSFORMERS_OFFLINE=1`, `HF_DATASETS_OFFLINE=1`, `HF_HUB_OFFLINE=1`, and `CHROMA_TELEMETRY=False`.
+- The LLM interaction is intended to occur to a local Ollama endpoint. `settings.py` sets `CHROMA_TELEMETRY=False`.
 
 ### Storage model
 - File-based: `user_data/pulse_history.json`, `user_data/user_profile.json`.
@@ -280,7 +280,7 @@
 ### Threat model
 - Protects against accidental network leakage and default external telemetry:
   - CHROMA_TELEMETRY disabled.
-  - HF/transformers offline env flags set.
+  - Uses local Ollama for embeddings.
 - Does not protect against:
   - Host compromise (malicious software or root access).
   - Misconfiguration where remote LLM endpoints are permitted (must be audited).
@@ -343,7 +343,7 @@
 
 ### Resource constraints
 - LLMs (MedGemma) require significant RAM/VRAM. Ollama hosting may require GPU or optimized CPU inference.
-- Sentencetransformer embedding uses CPU if no GPU; embedding large KBs has CPU and memory cost.
+- Ollama embedding requires Ollama to be running.
 
 ### Performance considerations
 - Caching:
@@ -417,16 +417,16 @@
 - The host device is trusted and not compromised.
 - Ollama and local model files are available locally or via a local host endpoint; the system is not configured to use remote LLM endpoints in the default repo configuration.
 - Disk-level encryption and OS security are the recommended mechanism for protecting data at rest unless the operator integrates additional encryption.
-- CHROMA_TELEMETRY and offline mode flags (`TRANSFORMERS_OFFLINE`, `HF_DATASETS_OFFLINE`, `HF_HUB_OFFLINE`) set in `settings.py` are respected by installed libraries.
+- CHROMA_TELEMETRY disabled in `settings.py`.
 
 ---
 
 ## Appendix
 
 ### Notable config values (in `settings.py`) 
 - `DB_PATH = 'user_data/user_med_db'`
-- `LLM_MODEL = 'MedAIBase/MedGemma1.5:4b'`
-- `EMBEDDING_MODEL = 'all-MiniLM-L6-v2'`
+- `LLM_MODEL = 'medgemma:27b'`
+- `EMBEDDING_MODEL = 'nomic-embed-text'`
 - `RAG_TOP_K = 2`
 - Cache TTLs: contextualized_queries 300s, rag_cache 600s, user_context_cache 180s.
 

pyproject.toml

@@ -41,7 +41,7 @@ dependencies = [
     "chromadb>=0.4",
     "numpy>=1.24",
     "scipy>=1.10",
-    "sentence-transformers>=2.0",
+    "ollama>=0.3.0",
     "xhtml2pdf>=0.2.17",
     "markdown>=3.5",
     "pymupdf>=1.23",

requirements.txt

@@ -3,7 +3,7 @@ requests>=2.28
 chromadb>=0.4
 numpy>=1.24
 scipy>=1.10
-sentence-transformers>=2.0
+ollama>=0.3.0
 xhtml2pdf>=0.2.17
 markdown>=3.5
 pymupdf>=1.23

scripts/setup_project.sh

@@ -22,9 +22,12 @@ except ImportError:
     sys.exit("Python venv module missing. Install python3-venv package.")
 PY
 
-# 2. Pull MedGemma
+# 2. Pull Models
 echo "Pulling MedGemma model (this may take a few minutes)..."
-ollama pull MedAIBase/MedGemma1.5:4b
+ollama pull medgemma:27b
+
+echo "Pulling embedding model..."
+ollama pull nomic-embed-text
 
 # Verify Ollama service responds (helpful if serve is not running)
 if command -v curl >/dev/null 2>&1; then

src/selene/settings.py

@@ -8,11 +8,10 @@
 
 import os
 
-# Force libraries to only look for local files - must be set before
-# importing transformers, sentence_transformers, or huggingface_hub.
-os.environ.setdefault("TRANSFORMERS_OFFLINE", "1")
-os.environ.setdefault("HF_DATASETS_OFFLINE", "1")
-os.environ.setdefault("HF_HUB_OFFLINE", "1")
+# Disable offline HF flags since we no longer use Hugging Face for embeddings
+# os.environ.setdefault("TRANSFORMERS_OFFLINE", "1")
+# os.environ.setdefault("HF_DATASETS_OFFLINE", "1")
+# os.environ.setdefault("HF_HUB_OFFLINE", "1")
 from pathlib import Path
 
 # ============================================================================
@@ -39,7 +38,7 @@
 DB_PATH = str(USER_DATA_DIR / "user_med_db")
 MEDICAL_DOCS_COLLECTION = "medical_docs"
 CHAT_HISTORY_COLLECTION = "chat_history"
-EMBEDDING_MODEL = "all-MiniLM-L6-v2"
+EMBEDDING_MODEL = "nomic-embed-text"
 CHROMA_TELEMETRY = False
 
 # ============================================================================
@@ -48,7 +47,7 @@
 
 OLLAMA_BASE_URL = "http://localhost:11434"
 OLLAMA_TIMEOUT = 60
-LLM_MODEL = "MedAIBase/MedGemma1.5:4b"
+LLM_MODEL = "medgemma:27b"
 
 # ============================================================================
 # RAG & Chat History Retrieval
@@ -101,16 +100,17 @@
 def get_embedding_function():
     """Return a cached ChromaDB-compatible embedding function.
 
-    Uses ChromaDB's built-in ``SentenceTransformerEmbeddingFunction`` so the
-    persisted collection metadata stays consistent (type ``sentence_transformer``).
+    Uses ChromaDB's built-in ``OllamaEmbeddingFunction`` so the
+    persisted collection metadata stays consistent.
     The model is loaded once and reused across all callers (Streamlit app and
     CLI tools like update_kb_chroma.py).
     """
     global _embedding_function_instance
     if _embedding_function_instance is None:
-        from chromadb.utils.embedding_functions import SentenceTransformerEmbeddingFunction
+        from chromadb.utils.embedding_functions import OllamaEmbeddingFunction
 
-        _embedding_function_instance = SentenceTransformerEmbeddingFunction(
+        _embedding_function_instance = OllamaEmbeddingFunction(
+            url=f"{OLLAMA_BASE_URL}/api/embeddings",
             model_name=EMBEDDING_MODEL
         )
     return _embedding_function_instance

src/selene/storage/chat_db.py

@@ -45,7 +45,7 @@ class ChatDBConfig:
 def _get_chat_client():
     """
     Returns a (collection, None) tuple on success, cached for the app lifetime.
-    Uses the same SentenceTransformer embedding model as med_logic so both
+    Uses the same Ollama embedding model as med_logic so both
     collections live in the same vector space - consistent and efficient.
     """
     try: