feat: LiteLLM model settings, dynamic discovery, and UI selectors

README.md

@@ -94,7 +94,7 @@ A high-performance, multi-repo Model Context Protocol (MCP) server providing **s
   - **Local Storage**: Managed file explorer, direct file upload modal with folder categorization, and file preview/replacement.
   - **Ingestion Catalog**: Unified multi-source explorer with source type filters, repository lookup, and file listings.
   - **Search & Inspector**: Interactive live hybrid search tester with RRF score previews, target type toggle (Code vs Docs), and syntax highlighted results.
-  - **Settings**: Vector Database manager (pgvector, Qdrant, & ChromaDB switcher & connection tester), multi-provider token cards, GitHub rate limit monitor, and interactive Custom Git Host Credential Vault table/modal.
+  - **Settings**: Vector Database manager (pgvector, Qdrant, & ChromaDB switcher & connection tester), LiteLLM Model Discovery with dynamic categorized model dropdowns (Embeddings, Vision OCR, and Chat models), multi-provider token cards, GitHub rate limit monitor, and interactive Custom Git Host Credential Vault table/modal.
   - **Diagnostics & Logs**: Real-time log viewer with level filtering (ALL, INFO, WARNING, ERROR, DEBUG), keyword search, traceback modal/drawer, and buffer clearing.
 
 ---

REQUIREMENTS.md

@@ -2,7 +2,7 @@
 
 > **Note:** This document is automatically generated and verified against the live test suite by `scripts/generate_requirements.py` and `tests/backend/test_requirements_sync.py`.
 
-**Test Verification Baseline:** **905 Automated Tests** (598 Pytest Backend + 261 Vitest Frontend + 46 Playwright E2E).
+**Test Verification Baseline:** **923 Automated Tests** (611 Pytest Backend + 266 Vitest Frontend + 46 Playwright E2E).
 
 ---
 
@@ -949,6 +949,15 @@ persisting all records and vector points correctly across multiple flushes._
 - `test_incremental_pipeline_clone_error_resilience` - _Verifies that a failure during shallow clone records an error in git_repositories
 and leaves the prior indexed state intact without data loss._
 
+#### `tests/test_litellm_service.py` (7 tests)
+- `test_discover_models_success`
+- `test_discover_models_timeout`
+- `test_discover_models_connect_error`
+- `test_discover_models_http_401_unauthorized`
+- `test_discover_models_http_500_error`
+- `test_discover_models_url_normalization`
+- `test_discover_models_default_resolution`
+
 #### `tests/test_local_storage_indexing.py` (4 tests)
 - `test_incremental_indexing_on_save`
 - `test_incremental_indexing_code_file`
@@ -980,6 +989,16 @@ and leaves the prior indexed state intact without data loss._
 - `test_what_is_ingested_detailed_with_data`
 - `test_tool_registration`
 
+#### `tests/test_model_discovery_api.py` (2 tests)
+- `test_api_discover_models_endpoint`
+- `test_api_embedding_settings_get_and_post_with_models`
+
+#### `tests/test_model_metadata_persistence.py` (4 tests)
+- `test_embedding_db_config_defaults`
+- `test_env_variable_fallbacks`
+- `test_set_embedding_db_config_persists_to_system_metadata`
+- `test_update_embedding_config_service`
+
 #### `tests/test_pdf_extractor.py` (8 tests)
 - `test_extract_digital_pdf_text`
 - `test_extract_scanned_pdf_triggers_vision_ocr`
@@ -1138,11 +1157,16 @@ and leaves the prior indexed state intact without data loss._
 - displays error toast when log fetching fails
 - renders responsive layout elements for toolbar, search input, and log entry stream
 
-#### `EmbeddingSettings.test.tsx` (4 tests)
+#### `EmbeddingSettings.test.tsx` (9 tests)
 - renders loading state when embedding configuration is not yet loaded
 - renders active status with hardware metrics and local model parameters
 - handles provider switch to API and updates form fields
 - handles changes to CPU threads and batch size
+- renders model discovery controls and triggers onDiscoverModels when button clicked
+- renders discovered model dropdowns and allows selecting models
+- displays discovery error banner when LiteLLM is unreachable
+- switches to manual input when Custom is selected from dropdown or link clicked
+- disables discover button and shows spinner while isDiscovering is true
 
 #### `GitRepoManager.test.tsx` (13 tests)
 - renders repository list with status badges, auto-sync buttons, and details

app/api/routers/settings.py

@@ -18,6 +18,7 @@
 import app.services.vector_store as vs_service
 import app.services.indexing as idx_service
 import app.services.embeddings as emb_service
+import app.services.litellm_service as litellm_service
 
 logger = logging.getLogger("contextcortex.api")
 
@@ -116,6 +117,8 @@ def _count(table):
             "embedding_provider": emb_cfg["provider"],
             "dense_model": emb_cfg["dense_model"],
             "sparse_model": emb_cfg["sparse_model"],
+            "vision_ocr_model": emb_cfg.get("vision_ocr_model"),
+            "chat_model": emb_cfg.get("chat_model"),
             "embedding_threads": emb_cfg["threads"],
             "embedding_batch_size": emb_cfg["batch_size"],
             "system_cpus": emb_cfg.get("system_cpus", 2),
@@ -319,6 +322,15 @@ def _reindex():
         logger.error(f"Error switching vector store backend: {e}")
         return JSONResponse(status_code=500, content={"status": "error", "error": str(e), "message": str(e)})
 
+@router.get("/admin/api/models/discover")
+async def api_discover_models(url: Optional[str] = None, api_key: Optional[str] = None):
+    try:
+        res = await litellm_service.discover_models(url=url, api_key=api_key)
+        return res
+    except Exception as e:
+        logger.error(f"Error discovering models: {e}")
+        return JSONResponse(status_code=500, content={"status": "error", "error": str(e), "message": str(e)})
+
 @router.get("/admin/api/settings/embedding")
 async def api_get_embedding_settings():
     try:
@@ -339,6 +351,8 @@ async def api_save_embedding_settings(payload: EmbeddingSettingsRequest):
             batch_size=payload.batch_size,
             litellm_url=payload.litellm_url,
             litellm_api_key=payload.litellm_api_key,
+            vision_ocr_model=payload.vision_ocr_model,
+            chat_model=payload.chat_model,
         )
         return {
             "status": "success",

app/models/schemas.py

@@ -170,6 +170,8 @@ class EmbeddingSettingsRequest(BaseModel):
     batch_size: Optional[int] = Field(default=None, ge=1, le=1024)
     litellm_url: Optional[str] = None
     litellm_api_key: Optional[str] = None
+    vision_ocr_model: Optional[str] = None
+    chat_model: Optional[str] = None
 
 # Topology Graph Models
 class TopologyNode(BaseModel):

app/services/database/__init__.py

@@ -36,6 +36,8 @@
     detect_system_resources,
     get_embedding_db_config,
     set_embedding_db_config,
+    get_vision_ocr_model,
+    get_chat_model,
 )
 from app.services.database.credentials import (
     list_git_host_credentials,
@@ -101,6 +103,8 @@
     "detect_system_resources",
     "get_embedding_db_config",
     "set_embedding_db_config",
+    "get_vision_ocr_model",
+    "get_chat_model",
     "list_git_host_credentials",
     "get_git_host_credential",
     "save_git_host_credential",

app/services/database/connection.py

@@ -224,6 +224,9 @@ def _resolve_default_embedding_config(conn: Optional[Any] = None) -> Dict[str, A
     litellm_url = os.getenv("LITELLM_URL", "http://litellm:4000/v1").strip()
     litellm_api_key = os.getenv("LITELLM_API_KEY", "dummy").strip()
 
+    vision_ocr_model = os.getenv("VISION_OCR_MODEL", "gemini-2.5-flash").strip()
+    chat_model = os.getenv("CHAT_MODEL", "gemini-2.5-flash").strip()
+
     return {
         "provider": provider,
         "dense_model": dense_model,
@@ -232,6 +235,8 @@ def _resolve_default_embedding_config(conn: Optional[Any] = None) -> Dict[str, A
         "batch_size": max(1, batch_size),
         "litellm_url": litellm_url,
         "litellm_api_key": litellm_api_key,
+        "vision_ocr_model": vision_ocr_model,
+        "chat_model": chat_model,
         "system_cpus": sys_res["cpus"],
         "system_memory_gb": sys_res["memory_gb"],
     }
@@ -245,6 +250,8 @@ def get_embedding_db_config() -> Dict[str, Any]:
     batch_size_str = get_metadata("embedding_batch_size")
     litellm_url = get_metadata("embedding_litellm_url")
     litellm_api_key = get_metadata("embedding_litellm_api_key")
+    vision_ocr_model = get_metadata("vision_ocr_model") or get_metadata("embedding_vision_ocr_model")
+    chat_model = get_metadata("chat_model") or get_metadata("embedding_chat_model")
 
     default_cfg = _resolve_default_embedding_config()
 
@@ -259,6 +266,8 @@ def get_embedding_db_config() -> Dict[str, Any]:
         "batch_size": max(1, batch_size),
         "litellm_url": (litellm_url or default_cfg["litellm_url"]).strip(),
         "litellm_api_key": litellm_api_key or default_cfg["litellm_api_key"],
+        "vision_ocr_model": (vision_ocr_model or default_cfg["vision_ocr_model"]).strip(),
+        "chat_model": (chat_model or default_cfg["chat_model"]).strip(),
         "system_cpus": default_cfg["system_cpus"],
         "system_memory_gb": default_cfg["system_memory_gb"],
     }
@@ -272,6 +281,8 @@ def set_embedding_db_config(
     batch_size: Optional[int] = None,
     litellm_url: Optional[str] = None,
     litellm_api_key: Optional[str] = None,
+    vision_ocr_model: Optional[str] = None,
+    chat_model: Optional[str] = None,
 ):
     if provider is not None:
         set_metadata("embedding_provider", provider.lower().strip())
@@ -287,3 +298,25 @@ def set_embedding_db_config(
         set_metadata("embedding_litellm_url", litellm_url.strip())
     if litellm_api_key is not None:
         set_metadata("embedding_litellm_api_key", litellm_api_key.strip())
+    if vision_ocr_model is not None:
+        set_metadata("vision_ocr_model", vision_ocr_model.strip())
+        set_metadata("embedding_vision_ocr_model", vision_ocr_model.strip())
+    if chat_model is not None:
+        set_metadata("chat_model", chat_model.strip())
+        set_metadata("embedding_chat_model", chat_model.strip())
+
+
+def get_vision_ocr_model() -> str:
+    """Returns stored vision OCR model or fallback to env/default."""
+    stored = get_metadata("vision_ocr_model") or get_metadata("embedding_vision_ocr_model")
+    if stored and stored.strip():
+        return stored.strip()
+    return (os.getenv("VISION_OCR_MODEL") or "gemini-2.5-flash").strip()
+
+
+def get_chat_model() -> str:
+    """Returns stored chat model or fallback to env/default."""
+    stored = get_metadata("chat_model") or get_metadata("embedding_chat_model")
+    if stored and stored.strip():
+        return stored.strip()
+    return (os.getenv("CHAT_MODEL") or "gemini-2.5-flash").strip()

app/services/embeddings.py

@@ -23,6 +23,8 @@ def _get_db_emb_config() -> Dict[str, Any]:
 EMBEDDING_BATCH_SIZE = _initial_cfg["batch_size"]
 LITELLM_URL = _initial_cfg["litellm_url"]
 LITELLM_API_KEY = _initial_cfg.get("litellm_api_key", os.getenv("LITELLM_API_KEY", "dummy"))
+VISION_OCR_MODEL = _initial_cfg.get("vision_ocr_model", os.getenv("VISION_OCR_MODEL", "gemini-2.5-flash"))
+CHAT_MODEL = _initial_cfg.get("chat_model", os.getenv("CHAT_MODEL", "gemini-2.5-flash"))
 
 _dense_model = None
 _sparse_model = None
@@ -38,10 +40,13 @@ def init_embeddings(
     batch_size: Optional[int] = None,
     litellm_url: Optional[str] = None,
     litellm_api_key: Optional[str] = None,
+    vision_ocr_model: Optional[str] = None,
+    chat_model: Optional[str] = None,
 ):
     global _dense_model, _sparse_model, _openai_client
     global EMBEDDING_PROVIDER, DENSE_MODEL_NAME, SPARSE_MODEL_NAME
     global EMBEDDING_NUM_THREADS, EMBEDDING_BATCH_SIZE, LITELLM_URL, LITELLM_API_KEY
+    global VISION_OCR_MODEL, CHAT_MODEL
 
     cfg = _get_db_emb_config()
 
@@ -52,6 +57,8 @@ def init_embeddings(
     EMBEDDING_BATCH_SIZE = int(batch_size if batch_size is not None else cfg.get("batch_size", 32))
     LITELLM_URL = (litellm_url or cfg.get("litellm_url") or "http://litellm:4000/v1").strip()
     LITELLM_API_KEY = (litellm_api_key or cfg.get("litellm_api_key") or "dummy").strip()
+    VISION_OCR_MODEL = (vision_ocr_model or cfg.get("vision_ocr_model") or os.getenv("VISION_OCR_MODEL", "gemini-2.5-flash")).strip()
+    CHAT_MODEL = (chat_model or cfg.get("chat_model") or os.getenv("CHAT_MODEL", "gemini-2.5-flash")).strip()
 
     # Set underlying OpenMP / BLAS thread guard
     os.environ["OMP_NUM_THREADS"] = str(EMBEDDING_NUM_THREADS)
@@ -103,6 +110,8 @@ def get_embedding_config() -> Dict[str, Any]:
         "threads": EMBEDDING_NUM_THREADS,
         "batch_size": EMBEDDING_BATCH_SIZE,
         "litellm_url": LITELLM_URL,
+        "vision_ocr_model": VISION_OCR_MODEL,
+        "chat_model": CHAT_MODEL,
         "system_cpus": sys_res["cpus"],
         "system_memory_gb": sys_res["memory_gb"],
     }
@@ -115,6 +124,8 @@ def update_embedding_config(
     batch_size: Optional[int] = None,
     litellm_url: Optional[str] = None,
     litellm_api_key: Optional[str] = None,
+    vision_ocr_model: Optional[str] = None,
+    chat_model: Optional[str] = None,
 ) -> Dict[str, Any]:
     """Updates embedding configuration in SQLite and hot-reloads models in memory."""
     from app.services.database import set_embedding_db_config
@@ -126,6 +137,8 @@ def update_embedding_config(
         batch_size=batch_size,
         litellm_url=litellm_url,
         litellm_api_key=litellm_api_key,
+        vision_ocr_model=vision_ocr_model,
+        chat_model=chat_model,
     )
     init_embeddings(
         provider=provider,
@@ -135,6 +148,8 @@ def update_embedding_config(
         batch_size=batch_size,
         litellm_url=litellm_url,
         litellm_api_key=litellm_api_key,
+        vision_ocr_model=vision_ocr_model,
+        chat_model=chat_model,
     )
     return get_embedding_config()