hindsight-api 0.3.0__py3-none-any.whl → 0.4.1__py3-none-any.whl

hindsight_api/engine/consolidation/prompts.py

@@ -0,0 +1,77 @@
+"""Prompts for the consolidation engine."""
+
+CONSOLIDATION_SYSTEM_PROMPT = """You are a memory consolidation system. Your job is to convert facts into durable knowledge (observations) and merge with existing knowledge when appropriate.
+
+You must output ONLY valid JSON with no markdown formatting, no code blocks, and no additional text.
+
+## EXTRACT DURABLE KNOWLEDGE, NOT EPHEMERAL STATE
+Facts often describe events or actions. Extract the DURABLE KNOWLEDGE implied by the fact, not the transient state.
+
+Examples of extracting durable knowledge:
+- "User moved to Room 203" -> "Room 203 exists" (location exists, not where user is now)
+- "User visited Acme Corp at Room 105" -> "Acme Corp is located in Room 105"
+- "User took the elevator to floor 3" -> "Floor 3 is accessible by elevator"
+- "User met Sarah at the lobby" -> "Sarah can be found at the lobby"
+
+DO NOT track current user position/state as knowledge - that changes constantly.
+DO track permanent facts learned from the user's actions.
+
+## PRESERVE SPECIFIC DETAILS
+Keep names, locations, numbers, and other specifics. Do NOT:
+- Abstract into general principles
+- Generate business insights
+- Make knowledge generic
+
+GOOD examples:
+- Fact: "John likes pizza" -> "John likes pizza"
+- Fact: "Alice works at Google" -> "Alice works at Google"
+
+BAD examples:
+- "John likes pizza" -> "Understanding dietary preferences helps..." (TOO ABSTRACT)
+- "User is at Room 203" -> "User is currently at Room 203" (EPHEMERAL STATE)
+
+## MERGE RULES (when comparing to existing observations):
+1. REDUNDANT: Same information worded differently → update existing
+2. CONTRADICTION: Opposite information about same topic → update with history (e.g., "used to X, now Y")
+3. UPDATE: New state replacing old state → update with history
+
+## CRITICAL RULES:
+- NEVER merge facts about DIFFERENT people
+- NEVER merge unrelated topics (food preferences vs work vs hobbies)
+- When merging contradictions, capture the CHANGE (before → after)
+- Keep observations focused on ONE specific topic per person
+- The "text" field MUST contain durable knowledge, not ephemeral state
+- Do NOT include "tags" in output - tags are handled automatically"""
+
+CONSOLIDATION_USER_PROMPT = """Analyze this new fact and consolidate into knowledge.
+{mission_section}
+NEW FACT: {fact_text}
+
+EXISTING OBSERVATIONS (JSON array with source memories and dates):
+{observations_text}
+
+Each observation includes:
+- id: unique identifier for updating
+- text: the observation content
+- proof_count: number of supporting memories
+- tags: visibility scope (handled automatically)
+- created_at/updated_at: when observation was created/modified
+- occurred_start/occurred_end: temporal range of source facts
+- source_memories: array of supporting facts with their text and dates
+
+Instructions:
+1. Extract DURABLE KNOWLEDGE from the new fact (not ephemeral state)
+2. Review source_memories in existing observations to understand evidence
+3. Check dates to detect contradictions or updates
+4. Compare with observations:
+   - Same topic → UPDATE with learning_id
+   - New topic → CREATE new observation
+   - Purely ephemeral → return []
+
+Output JSON array of actions:
+[
+  {{"action": "update", "learning_id": "uuid-from-observations", "text": "updated knowledge", "reason": "..."}},
+  {{"action": "create", "text": "new durable knowledge", "reason": "..."}}
+]
+
+Return [] if fact contains no durable knowledge."""

hindsight_api/engine/cross_encoder.py

@@ -20,6 +20,7 @@ from ..config import (
     DEFAULT_RERANKER_FLASHRANK_CACHE_DIR,
     DEFAULT_RERANKER_FLASHRANK_MODEL,
     DEFAULT_RERANKER_LITELLM_MODEL,
+    DEFAULT_RERANKER_LOCAL_FORCE_CPU,
     DEFAULT_RERANKER_LOCAL_MAX_CONCURRENT,
     DEFAULT_RERANKER_LOCAL_MODEL,
     DEFAULT_RERANKER_PROVIDER,
@@ -33,6 +34,7 @@ from ..config import (
     ENV_RERANKER_FLASHRANK_CACHE_DIR,
     ENV_RERANKER_FLASHRANK_MODEL,
     ENV_RERANKER_LITELLM_MODEL,
+    ENV_RERANKER_LOCAL_FORCE_CPU,
     ENV_RERANKER_LOCAL_MAX_CONCURRENT,
     ENV_RERANKER_LOCAL_MODEL,
     ENV_RERANKER_PROVIDER,
@@ -99,7 +101,7 @@ class LocalSTCrossEncoder(CrossEncoderModel):
     _executor: ThreadPoolExecutor | None = None
     _max_concurrent: int = 4  # Limit concurrent CPU-bound reranking calls
 
-    def __init__(self, model_name: str | None = None, max_concurrent: int = 4):
+    def __init__(self, model_name: str | None = None, max_concurrent: int = 4, force_cpu: bool = False):
         """
         Initialize local SentenceTransformers cross-encoder.
 
@@ -108,8 +110,11 @@ class LocalSTCrossEncoder(CrossEncoderModel):
                        Default: cross-encoder/ms-marco-MiniLM-L-6-v2
             max_concurrent: Maximum concurrent reranking calls (default: 2).
                            Higher values may cause CPU thrashing under load.
+            force_cpu: Force CPU mode (avoids MPS/XPC issues on macOS in daemon mode).
+                      Default: False
         """
         self.model_name = model_name or DEFAULT_RERANKER_LOCAL_MODEL
+        self.force_cpu = force_cpu
         self._model = None
         LocalSTCrossEncoder._max_concurrent = max_concurrent
 
@@ -130,13 +135,38 @@ class LocalSTCrossEncoder(CrossEncoderModel):
                 "Install it with: pip install sentence-transformers"
             )
 
-        # Note: We use CPU even when GPU/MPS is available because:
-        # 1. The reranker model (MiniLM) is tiny (~22M params)
-        # 2. Batch sizes are small (~100-200 pairs)
-        # 3. Data transfer overhead to GPU outweighs compute benefit
-        # 4. CPU inference is actually faster for this workload
         logger.info(f"Reranker: initializing local provider with model {self.model_name}")
-        self._model = CrossEncoder(self.model_name)
+
+        # Determine device based on hardware availability.
+        # We always set low_cpu_mem_usage=False to prevent lazy loading (meta tensors)
+        # which can cause issues when accelerate is installed but no GPU is available.
+        # Note: We do NOT use device_map because CrossEncoder internally calls .to(device)
+        # after loading, which conflicts with accelerate's device_map handling.
+        import torch
+
+        # Force CPU mode if configured (used in daemon mode to avoid MPS/XPC issues on macOS)
+        if self.force_cpu:
+            device = "cpu"
+            logger.info("Reranker: forcing CPU mode (HINDSIGHT_API_RERANKER_LOCAL_FORCE_CPU=1)")
+        else:
+            # Check for GPU (CUDA) or Apple Silicon (MPS)
+            # Wrap in try-except to gracefully handle any device detection issues
+            # (e.g., in CI environments or when PyTorch is built without GPU support)
+            device = "cpu"  # Default to CPU
+            try:
+                has_gpu = torch.cuda.is_available() or (
+                    hasattr(torch.backends, "mps") and torch.backends.mps.is_available()
+                )
+                if has_gpu:
+                    device = None  # Let sentence-transformers auto-detect GPU/MPS
+            except Exception as e:
+                logger.warning(f"Failed to detect GPU/MPS, falling back to CPU: {e}")
+
+        self._model = CrossEncoder(
+            self.model_name,
+            device=device,
+            model_kwargs={"low_cpu_mem_usage": False},
+        )
 
         # Initialize shared executor (limited workers naturally limits concurrency)
         if LocalSTCrossEncoder._executor is None:
@@ -148,11 +178,108 @@ class LocalSTCrossEncoder(CrossEncoderModel):
         else:
             logger.info("Reranker: local provider initialized (using existing executor)")
 
+    def _is_xpc_error(self, error: Exception) -> bool:
+        """
+        Check if an error is an XPC connection error (macOS daemon issue).
+
+        On macOS, long-running daemons can lose XPC connections to system services
+        when the process is idle for extended periods.
+        """
+        error_str = str(error).lower()
+        return "xpc_error_connection_invalid" in error_str or "xpc error" in error_str
+
+    def _reinitialize_model_sync(self) -> None:
+        """
+        Clear and reinitialize the cross-encoder model synchronously.
+
+        This is used to recover from XPC errors on macOS where the
+        PyTorch/MPS backend loses its connection to system services.
+        """
+        logger.warning(f"Reinitializing reranker model {self.model_name} due to backend error")
+
+        # Clear existing model
+        self._model = None
+
+        # Force garbage collection to free resources
+        import gc
+
+        import torch
+
+        gc.collect()
+
+        # If using CUDA/MPS, clear the cache
+        if torch.cuda.is_available():
+            torch.cuda.empty_cache()
+        elif hasattr(torch.backends, "mps") and torch.backends.mps.is_available():
+            try:
+                torch.mps.empty_cache()
+            except AttributeError:
+                pass  # Method might not exist in all PyTorch versions
+
+        # Reinitialize the model
+        try:
+            from sentence_transformers import CrossEncoder
+        except ImportError:
+            raise ImportError(
+                "sentence-transformers is required for LocalSTCrossEncoder. "
+                "Install it with: pip install sentence-transformers"
+            )
+
+        # Determine device based on hardware availability
+        if self.force_cpu:
+            device = "cpu"
+        else:
+            # Wrap in try-except to gracefully handle any device detection issues
+            device = "cpu"  # Default to CPU
+            try:
+                has_gpu = torch.cuda.is_available() or (
+                    hasattr(torch.backends, "mps") and torch.backends.mps.is_available()
+                )
+                if has_gpu:
+                    device = None  # Let sentence-transformers auto-detect GPU/MPS
+            except Exception as e:
+                logger.warning(f"Failed to detect GPU/MPS during reinit, falling back to CPU: {e}")
+
+        self._model = CrossEncoder(
+            self.model_name,
+            device=device,
+            model_kwargs={"low_cpu_mem_usage": False},
+        )
+
+        logger.info("Reranker: local provider reinitialized successfully")
+
+    def _predict_with_recovery(self, pairs: list[tuple[str, str]]) -> list[float]:
+        """
+        Predict with automatic recovery from XPC errors.
+
+        This runs synchronously in the thread pool.
+        """
+        max_retries = 1
+        for attempt in range(max_retries + 1):
+            try:
+                scores = self._model.predict(pairs, show_progress_bar=False)
+                return scores.tolist() if hasattr(scores, "tolist") else list(scores)
+            except Exception as e:
+                # Check if this is an XPC error (macOS daemon issue)
+                if self._is_xpc_error(e) and attempt < max_retries:
+                    logger.warning(f"XPC error detected in reranker (attempt {attempt + 1}): {e}")
+                    try:
+                        self._reinitialize_model_sync()
+                        logger.info("Reranker reinitialized successfully, retrying prediction")
+                        continue
+                    except Exception as reinit_error:
+                        logger.error(f"Failed to reinitialize reranker: {reinit_error}")
+                        raise Exception(f"Failed to recover from XPC error: {str(e)}")
+                else:
+                    # Not an XPC error or out of retries
+                    raise
+
     async def predict(self, pairs: list[tuple[str, str]]) -> list[float]:
         """
         Score query-document pairs for relevance.
 
         Uses a dedicated thread pool with limited workers to prevent CPU thrashing.
+        Automatically recovers from XPC errors on macOS by reinitializing the model.
 
         Args:
             pairs: List of (query, document) tuples to score
@@ -165,11 +292,11 @@ class LocalSTCrossEncoder(CrossEncoderModel):
 
         # Use dedicated executor - limited workers naturally limits concurrency
         loop = asyncio.get_event_loop()
-        scores = await loop.run_in_executor(
+        return await loop.run_in_executor(
             LocalSTCrossEncoder._executor,
-            lambda: self._model.predict(pairs, show_progress_bar=False),
+            self._predict_with_recovery,
+            pairs,
         )
-        return scores.tolist() if hasattr(scores, "tolist") else list(scores)
 
 
 class RemoteTEICrossEncoder(CrossEncoderModel):
@@ -768,29 +895,33 @@ class LiteLLMCrossEncoder(CrossEncoderModel):
 
 def create_cross_encoder_from_env() -> CrossEncoderModel:
     """
-    Create a CrossEncoderModel instance based on environment variables.
+    Create a CrossEncoderModel instance based on configuration.
 
-    See hindsight_api.config for environment variable names and defaults.
+    Reads configuration via get_config() to ensure consistency across the codebase.
 
     Returns:
         Configured CrossEncoderModel instance
     """
-    provider = os.environ.get(ENV_RERANKER_PROVIDER, DEFAULT_RERANKER_PROVIDER).lower()
+    from ..config import get_config
+
+    config = get_config()
+    provider = config.reranker_provider.lower()
 
     if provider == "tei":
-        url = os.environ.get(ENV_RERANKER_TEI_URL)
+        url = config.reranker_tei_url
         if not url:
             raise ValueError(f"{ENV_RERANKER_TEI_URL} is required when {ENV_RERANKER_PROVIDER} is 'tei'")
-        batch_size = int(os.environ.get(ENV_RERANKER_TEI_BATCH_SIZE, str(DEFAULT_RERANKER_TEI_BATCH_SIZE)))
-        max_concurrent = int(os.environ.get(ENV_RERANKER_TEI_MAX_CONCURRENT, str(DEFAULT_RERANKER_TEI_MAX_CONCURRENT)))
-        return RemoteTEICrossEncoder(base_url=url, batch_size=batch_size, max_concurrent=max_concurrent)
+        return RemoteTEICrossEncoder(
+            base_url=url,
+            batch_size=config.reranker_tei_batch_size,
+            max_concurrent=config.reranker_tei_max_concurrent,
+        )
     elif provider == "local":
-        model = os.environ.get(ENV_RERANKER_LOCAL_MODEL)
-        model_name = model or DEFAULT_RERANKER_LOCAL_MODEL
-        max_concurrent = int(
-            os.environ.get(ENV_RERANKER_LOCAL_MAX_CONCURRENT, str(DEFAULT_RERANKER_LOCAL_MAX_CONCURRENT))
+        return LocalSTCrossEncoder(
+            model_name=config.reranker_local_model,
+            max_concurrent=config.reranker_local_max_concurrent,
+            force_cpu=config.reranker_local_force_cpu,
         )
-        return LocalSTCrossEncoder(model_name=model_name, max_concurrent=max_concurrent)
     elif provider == "cohere":
         api_key = os.environ.get(ENV_COHERE_API_KEY)
         if not api_key:

hindsight_api/engine/directives/__init__.py

@@ -0,0 +1,5 @@
+"""Directives module for hard rules injected into prompts."""
+
+from .models import Directive
+
+__all__ = ["Directive"]

hindsight_api/engine/directives/models.py

@@ -0,0 +1,37 @@
+"""Pydantic models for directives."""
+
+from datetime import datetime, timezone
+from uuid import UUID
+
+from pydantic import BaseModel, Field
+
+
+class Directive(BaseModel):
+    """A directive is a hard rule injected into prompts.
+
+    Directives are user-defined rules that guide agent behavior. Unlike mental models
+    which are automatically consolidated from memories, directives are explicit
+    instructions that are always included in relevant prompts.
+
+    Examples:
+    - "Always respond in formal English"
+    - "Never share personal data with third parties"
+    - "Prefer conservative investment recommendations"
+    """
+
+    id: UUID = Field(description="Unique identifier")
+    bank_id: str = Field(description="Bank this directive belongs to")
+    name: str = Field(description="Human-readable name")
+    content: str = Field(description="The directive text to inject into prompts")
+    priority: int = Field(default=0, description="Higher priority directives are injected first")
+    is_active: bool = Field(default=True, description="Whether this directive is currently active")
+    tags: list[str] = Field(default_factory=list, description="Tags for filtering")
+    created_at: datetime = Field(
+        default_factory=lambda: datetime.now(timezone.utc), description="When this directive was created"
+    )
+    updated_at: datetime = Field(
+        default_factory=lambda: datetime.now(timezone.utc), description="When this directive was last updated"
+    )
+
+    class Config:
+        from_attributes = True

hindsight_api/engine/embeddings.py

@@ -18,6 +18,7 @@ import httpx
 from ..config import (
     DEFAULT_EMBEDDINGS_COHERE_MODEL,
     DEFAULT_EMBEDDINGS_LITELLM_MODEL,
+    DEFAULT_EMBEDDINGS_LOCAL_FORCE_CPU,
     DEFAULT_EMBEDDINGS_LOCAL_MODEL,
     DEFAULT_EMBEDDINGS_OPENAI_MODEL,
     DEFAULT_EMBEDDINGS_PROVIDER,
@@ -26,6 +27,7 @@ from ..config import (
     ENV_EMBEDDINGS_COHERE_BASE_URL,
     ENV_EMBEDDINGS_COHERE_MODEL,
     ENV_EMBEDDINGS_LITELLM_MODEL,
+    ENV_EMBEDDINGS_LOCAL_FORCE_CPU,
     ENV_EMBEDDINGS_LOCAL_MODEL,
     ENV_EMBEDDINGS_OPENAI_API_KEY,
     ENV_EMBEDDINGS_OPENAI_BASE_URL,
@@ -92,15 +94,18 @@ class LocalSTEmbeddings(Embeddings):
     The embedding dimension is auto-detected from the model.
     """
 
-    def __init__(self, model_name: str | None = None):
+    def __init__(self, model_name: str | None = None, force_cpu: bool = False):
         """
         Initialize local SentenceTransformers embeddings.
 
         Args:
             model_name: Name of the SentenceTransformer model to use.
                        Default: BAAI/bge-small-en-v1.5
+            force_cpu: Force CPU mode (avoids MPS/XPC issues on macOS in daemon mode).
+                      Default: False
         """
         self.model_name = model_name or DEFAULT_EMBEDDINGS_LOCAL_MODEL
+        self.force_cpu = force_cpu
         self._model = None
         self._dimension: int | None = None
 
@@ -128,20 +133,115 @@ class LocalSTEmbeddings(Embeddings):
             )
 
         logger.info(f"Embeddings: initializing local provider with model {self.model_name}")
-        # Disable lazy loading (meta tensors) which causes issues with newer transformers/accelerate
-        # Setting low_cpu_mem_usage=False and device_map=None ensures tensors are fully materialized
+
+        # Determine device based on hardware availability.
+        # We always set low_cpu_mem_usage=False to prevent lazy loading (meta tensors)
+        # which can cause issues when accelerate is installed but no GPU is available.
+        import torch
+
+        # Force CPU mode if configured (used in daemon mode to avoid MPS/XPC issues on macOS)
+        if self.force_cpu:
+            device = "cpu"
+            logger.info("Embeddings: forcing CPU mode")
+        else:
+            # Check for GPU (CUDA) or Apple Silicon (MPS)
+            # Wrap in try-except to gracefully handle any device detection issues
+            # (e.g., in CI environments or when PyTorch is built without GPU support)
+            device = "cpu"  # Default to CPU
+            try:
+                has_gpu = torch.cuda.is_available() or (
+                    hasattr(torch.backends, "mps") and torch.backends.mps.is_available()
+                )
+                if has_gpu:
+                    device = None  # Let sentence-transformers auto-detect GPU/MPS
+            except Exception as e:
+                logger.warning(f"Failed to detect GPU/MPS, falling back to CPU: {e}")
+
         self._model = SentenceTransformer(
             self.model_name,
-            model_kwargs={"low_cpu_mem_usage": False, "device_map": None},
+            device=device,
+            model_kwargs={"low_cpu_mem_usage": False},
         )
 
         self._dimension = self._model.get_sentence_embedding_dimension()
         logger.info(f"Embeddings: local provider initialized (dim: {self._dimension})")
 
+    def _is_xpc_error(self, error: Exception) -> bool:
+        """
+        Check if an error is an XPC connection error (macOS daemon issue).
+
+        On macOS, long-running daemons can lose XPC connections to system services
+        when the process is idle for extended periods.
+        """
+        error_str = str(error).lower()
+        return "xpc_error_connection_invalid" in error_str or "xpc error" in error_str
+
+    def _reinitialize_model_sync(self) -> None:
+        """
+        Clear and reinitialize the embedding model synchronously.
+
+        This is used to recover from XPC errors on macOS where the
+        PyTorch/MPS backend loses its connection to system services.
+        """
+        logger.warning(f"Reinitializing embedding model {self.model_name} due to backend error")
+
+        # Clear existing model
+        self._model = None
+
+        # Force garbage collection to free resources
+        import gc
+
+        import torch
+
+        gc.collect()
+
+        # If using CUDA/MPS, clear the cache
+        if torch.cuda.is_available():
+            torch.cuda.empty_cache()
+        elif hasattr(torch.backends, "mps") and torch.backends.mps.is_available():
+            try:
+                torch.mps.empty_cache()
+            except AttributeError:
+                pass  # Method might not exist in all PyTorch versions
+
+        # Reinitialize the model (inline version of initialize() but synchronous)
+        try:
+            from sentence_transformers import SentenceTransformer
+        except ImportError:
+            raise ImportError(
+                "sentence-transformers is required for LocalSTEmbeddings. "
+                "Install it with: pip install sentence-transformers"
+            )
+
+        # Determine device based on hardware availability
+        if self.force_cpu:
+            device = "cpu"
+        else:
+            # Wrap in try-except to gracefully handle any device detection issues
+            device = "cpu"  # Default to CPU
+            try:
+                has_gpu = torch.cuda.is_available() or (
+                    hasattr(torch.backends, "mps") and torch.backends.mps.is_available()
+                )
+                if has_gpu:
+                    device = None  # Let sentence-transformers auto-detect GPU/MPS
+            except Exception as e:
+                logger.warning(f"Failed to detect GPU/MPS during reinit, falling back to CPU: {e}")
+
+        self._model = SentenceTransformer(
+            self.model_name,
+            device=device,
+            model_kwargs={"low_cpu_mem_usage": False},
+        )
+
+        logger.info("Embeddings: local provider reinitialized successfully")
+
     def encode(self, texts: list[str]) -> list[list[float]]:
         """
         Generate embeddings for a list of texts.
 
+        Automatically recovers from XPC errors on macOS by reinitializing the model.
+
         Args:
             texts: List of text strings to encode
 
@@ -150,8 +250,27 @@ class LocalSTEmbeddings(Embeddings):
         """
         if self._model is None:
             raise RuntimeError("Embeddings not initialized. Call initialize() first.")
-        embeddings = self._model.encode(texts, convert_to_numpy=True, show_progress_bar=False)
-        return [emb.tolist() for emb in embeddings]
+
+        # Try encoding with automatic recovery from XPC errors
+        max_retries = 1
+        for attempt in range(max_retries + 1):
+            try:
+                embeddings = self._model.encode(texts, convert_to_numpy=True, show_progress_bar=False)
+                return [emb.tolist() for emb in embeddings]
+            except Exception as e:
+                # Check if this is an XPC error (macOS daemon issue)
+                if self._is_xpc_error(e) and attempt < max_retries:
+                    logger.warning(f"XPC error detected in embedding generation (attempt {attempt + 1}): {e}")
+                    try:
+                        self._reinitialize_model_sync()
+                        logger.info("Model reinitialized successfully, retrying embedding generation")
+                        continue
+                    except Exception as reinit_error:
+                        logger.error(f"Failed to reinitialize model: {reinit_error}")
+                        raise Exception(f"Failed to recover from XPC error: {str(e)}")
+                else:
+                    # Not an XPC error or out of retries
+                    raise
 
 
 class RemoteTEIEmbeddings(Embeddings):
@@ -673,24 +792,28 @@ class LiteLLMEmbeddings(Embeddings):
 
 def create_embeddings_from_env() -> Embeddings:
     """
-    Create an Embeddings instance based on environment variables.
+    Create an Embeddings instance based on configuration.
 
-    See hindsight_api.config for environment variable names and defaults.
+    Reads configuration via get_config() to ensure consistency across the codebase.
 
     Returns:
         Configured Embeddings instance
     """
-    provider = os.environ.get(ENV_EMBEDDINGS_PROVIDER, DEFAULT_EMBEDDINGS_PROVIDER).lower()
+    from ..config import get_config
+
+    config = get_config()
+    provider = config.embeddings_provider.lower()
 
     if provider == "tei":
-        url = os.environ.get(ENV_EMBEDDINGS_TEI_URL)
+        url = config.embeddings_tei_url
         if not url:
             raise ValueError(f"{ENV_EMBEDDINGS_TEI_URL} is required when {ENV_EMBEDDINGS_PROVIDER} is 'tei'")
         return RemoteTEIEmbeddings(base_url=url)
     elif provider == "local":
-        model = os.environ.get(ENV_EMBEDDINGS_LOCAL_MODEL)
-        model_name = model or DEFAULT_EMBEDDINGS_LOCAL_MODEL
-        return LocalSTEmbeddings(model_name=model_name)
+        return LocalSTEmbeddings(
+            model_name=config.embeddings_local_model,
+            force_cpu=config.embeddings_local_force_cpu,
+        )
     elif provider == "openai":
         # Use dedicated embeddings API key, or fall back to LLM API key
         api_key = os.environ.get(ENV_EMBEDDINGS_OPENAI_API_KEY) or os.environ.get(ENV_LLM_API_KEY)