hindsight-api 0.4.0__tar.gz → 0.4.1__tar.gz

{hindsight_api-0.4.0 → hindsight_api-0.4.1}/.gitignore

@@ -45,9 +45,12 @@ hindsight-docs/static/llms-full.txt
 
 hindsight-dev/benchmarks/locomo/results/
 hindsight-dev/benchmarks/longmemeval/results/
+hindsight-dev/benchmarks/consolidation/results/
+benchmarks/results/
 hindsight-cli/target
 hindsight-clients/rust/target
 .claude
 whats-next.md
 TASK.md
-CHANGELOG.md
+# Changelog is now tracked in hindsight-docs/src/pages/changelog.md
+# CHANGELOG.md

{hindsight_api-0.4.0 → hindsight_api-0.4.1}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: hindsight-api
-Version: 0.4.0
+Version: 0.4.1
 Summary: Hindsight: Agent Memory That Works Like Human Memory
 Requires-Python: >=3.11
 Requires-Dist: aiohttp>=3.13.3

{hindsight_api-0.4.0 → hindsight_api-0.4.1}/hindsight_api/__init__.py

@@ -46,4 +46,4 @@ __all__ = [
     "RemoteTEICrossEncoder",
     "LLMConfig",
 ]
-__version__ = "0.1.0"
+__version__ = "0.4.1"

{hindsight_api-0.4.0 → hindsight_api-0.4.1}/hindsight_api/api/http.py

@@ -1323,7 +1323,7 @@ class VersionResponse(BaseModel):
     model_config = ConfigDict(
         json_schema_extra={
             "example": {
-                "api_version": "1.0.0",
+                "api_version": "0.4.0",
                 "features": {
                     "observations": False,
                     "mcp": True,
@@ -1567,11 +1567,12 @@ def _register_routes(app: FastAPI):
         Returns version info and feature flags that can be used by clients
         to determine which capabilities are available.
         """
+        from hindsight_api import __version__
         from hindsight_api.config import get_config
 
         config = get_config()
         return VersionResponse(
-            api_version="1.0.0",
+            api_version=__version__,
             features=FeaturesInfo(
                 observations=config.enable_observations,
                 mcp=config.mcp_enabled,

{hindsight_api-0.4.0 → hindsight_api-0.4.1}/hindsight_api/config.py

@@ -20,6 +20,7 @@ logger = logging.getLogger(__name__)
 
 # Environment variable names
 ENV_DATABASE_URL = "HINDSIGHT_API_DATABASE_URL"
+ENV_DATABASE_SCHEMA = "HINDSIGHT_API_DATABASE_SCHEMA"
 ENV_LLM_PROVIDER = "HINDSIGHT_API_LLM_PROVIDER"
 ENV_LLM_API_KEY = "HINDSIGHT_API_LLM_API_KEY"
 ENV_LLM_MODEL = "HINDSIGHT_API_LLM_MODEL"
@@ -46,6 +47,7 @@ ENV_CONSOLIDATION_LLM_BASE_URL = "HINDSIGHT_API_CONSOLIDATION_LLM_BASE_URL"
 
 ENV_EMBEDDINGS_PROVIDER = "HINDSIGHT_API_EMBEDDINGS_PROVIDER"
 ENV_EMBEDDINGS_LOCAL_MODEL = "HINDSIGHT_API_EMBEDDINGS_LOCAL_MODEL"
+ENV_EMBEDDINGS_LOCAL_FORCE_CPU = "HINDSIGHT_API_EMBEDDINGS_LOCAL_FORCE_CPU"
 ENV_EMBEDDINGS_TEI_URL = "HINDSIGHT_API_EMBEDDINGS_TEI_URL"
 ENV_EMBEDDINGS_OPENAI_API_KEY = "HINDSIGHT_API_EMBEDDINGS_OPENAI_API_KEY"
 ENV_EMBEDDINGS_OPENAI_MODEL = "HINDSIGHT_API_EMBEDDINGS_OPENAI_MODEL"
@@ -65,6 +67,7 @@ ENV_RERANKER_LITELLM_MODEL = "HINDSIGHT_API_RERANKER_LITELLM_MODEL"
 
 ENV_RERANKER_PROVIDER = "HINDSIGHT_API_RERANKER_PROVIDER"
 ENV_RERANKER_LOCAL_MODEL = "HINDSIGHT_API_RERANKER_LOCAL_MODEL"
+ENV_RERANKER_LOCAL_FORCE_CPU = "HINDSIGHT_API_RERANKER_LOCAL_FORCE_CPU"
 ENV_RERANKER_LOCAL_MAX_CONCURRENT = "HINDSIGHT_API_RERANKER_LOCAL_MAX_CONCURRENT"
 ENV_RERANKER_TEI_URL = "HINDSIGHT_API_RERANKER_TEI_URL"
 ENV_RERANKER_TEI_BATCH_SIZE = "HINDSIGHT_API_RERANKER_TEI_BATCH_SIZE"
@@ -98,6 +101,7 @@ ENV_RETAIN_OBSERVATIONS_ASYNC = "HINDSIGHT_API_RETAIN_OBSERVATIONS_ASYNC"
 # Observations settings (consolidated knowledge from facts)
 ENV_ENABLE_OBSERVATIONS = "HINDSIGHT_API_ENABLE_OBSERVATIONS"
 ENV_CONSOLIDATION_BATCH_SIZE = "HINDSIGHT_API_CONSOLIDATION_BATCH_SIZE"
+ENV_CONSOLIDATION_MAX_TOKENS = "HINDSIGHT_API_CONSOLIDATION_MAX_TOKENS"
 
 # Optimization flags
 ENV_SKIP_LLM_VERIFICATION = "HINDSIGHT_API_SKIP_LLM_VERIFICATION"
@@ -125,6 +129,7 @@ ENV_REFLECT_MAX_ITERATIONS = "HINDSIGHT_API_REFLECT_MAX_ITERATIONS"
 
 # Default values
 DEFAULT_DATABASE_URL = "pg0"
+DEFAULT_DATABASE_SCHEMA = "public"
 DEFAULT_LLM_PROVIDER = "openai"
 DEFAULT_LLM_MODEL = "gpt-5-mini"
 DEFAULT_LLM_MAX_CONCURRENT = 32
@@ -132,11 +137,13 @@ DEFAULT_LLM_TIMEOUT = 120.0  # seconds
 
 DEFAULT_EMBEDDINGS_PROVIDER = "local"
 DEFAULT_EMBEDDINGS_LOCAL_MODEL = "BAAI/bge-small-en-v1.5"
+DEFAULT_EMBEDDINGS_LOCAL_FORCE_CPU = False  # Force CPU mode for local embeddings (avoids MPS/XPC issues on macOS)
 DEFAULT_EMBEDDINGS_OPENAI_MODEL = "text-embedding-3-small"
 DEFAULT_EMBEDDING_DIMENSION = 384
 
 DEFAULT_RERANKER_PROVIDER = "local"
 DEFAULT_RERANKER_LOCAL_MODEL = "cross-encoder/ms-marco-MiniLM-L-6-v2"
+DEFAULT_RERANKER_LOCAL_FORCE_CPU = False  # Force CPU mode for local reranker (avoids MPS/XPC issues on macOS)
 DEFAULT_RERANKER_LOCAL_MAX_CONCURRENT = 4  # Limit concurrent CPU-bound reranking to prevent thrashing
 DEFAULT_RERANKER_TEI_BATCH_SIZE = 128
 DEFAULT_RERANKER_TEI_MAX_CONCURRENT = 8
@@ -177,6 +184,7 @@ DEFAULT_RETAIN_OBSERVATIONS_ASYNC = False  # Run observation generation async (a
 # Observations defaults (consolidated knowledge from facts)
 DEFAULT_ENABLE_OBSERVATIONS = True  # Observations enabled by default
 DEFAULT_CONSOLIDATION_BATCH_SIZE = 50  # Memories to load per batch (internal memory optimization)
+DEFAULT_CONSOLIDATION_MAX_TOKENS = 1024  # Max tokens for recall when finding related observations
 
 # Database migrations
 DEFAULT_RUN_MIGRATIONS_ON_STARTUP = True
@@ -270,6 +278,7 @@ class HindsightConfig:
 
     # Database
     database_url: str
+    database_schema: str
 
     # LLM (default, used as fallback for per-operation config)
     llm_provider: str
@@ -298,6 +307,7 @@ class HindsightConfig:
     # Embeddings
     embeddings_provider: str
     embeddings_local_model: str
+    embeddings_local_force_cpu: bool
     embeddings_tei_url: str | None
     embeddings_openai_base_url: str | None
     embeddings_cohere_base_url: str | None
@@ -305,6 +315,8 @@ class HindsightConfig:
     # Reranker
     reranker_provider: str
     reranker_local_model: str
+    reranker_local_force_cpu: bool
+    reranker_local_max_concurrent: int
     reranker_tei_url: str | None
     reranker_tei_batch_size: int
     reranker_tei_max_concurrent: int
@@ -336,6 +348,7 @@ class HindsightConfig:
     # Observations settings (consolidated knowledge from facts)
     enable_observations: bool
     consolidation_batch_size: int
+    consolidation_max_tokens: int
 
     # Optimization flags
     skip_llm_verification: bool
@@ -367,6 +380,7 @@ class HindsightConfig:
         return cls(
             # Database
             database_url=os.getenv(ENV_DATABASE_URL, DEFAULT_DATABASE_URL),
+            database_schema=os.getenv(ENV_DATABASE_SCHEMA, DEFAULT_DATABASE_SCHEMA),
             # LLM
             llm_provider=os.getenv(ENV_LLM_PROVIDER, DEFAULT_LLM_PROVIDER),
             llm_api_key=os.getenv(ENV_LLM_API_KEY),
@@ -390,12 +404,23 @@ class HindsightConfig:
             # Embeddings
             embeddings_provider=os.getenv(ENV_EMBEDDINGS_PROVIDER, DEFAULT_EMBEDDINGS_PROVIDER),
             embeddings_local_model=os.getenv(ENV_EMBEDDINGS_LOCAL_MODEL, DEFAULT_EMBEDDINGS_LOCAL_MODEL),
+            embeddings_local_force_cpu=os.getenv(
+                ENV_EMBEDDINGS_LOCAL_FORCE_CPU, str(DEFAULT_EMBEDDINGS_LOCAL_FORCE_CPU)
+            ).lower()
+            in ("true", "1"),
             embeddings_tei_url=os.getenv(ENV_EMBEDDINGS_TEI_URL),
             embeddings_openai_base_url=os.getenv(ENV_EMBEDDINGS_OPENAI_BASE_URL) or None,
             embeddings_cohere_base_url=os.getenv(ENV_EMBEDDINGS_COHERE_BASE_URL) or None,
             # Reranker
             reranker_provider=os.getenv(ENV_RERANKER_PROVIDER, DEFAULT_RERANKER_PROVIDER),
             reranker_local_model=os.getenv(ENV_RERANKER_LOCAL_MODEL, DEFAULT_RERANKER_LOCAL_MODEL),
+            reranker_local_force_cpu=os.getenv(
+                ENV_RERANKER_LOCAL_FORCE_CPU, str(DEFAULT_RERANKER_LOCAL_FORCE_CPU)
+            ).lower()
+            in ("true", "1"),
+            reranker_local_max_concurrent=int(
+                os.getenv(ENV_RERANKER_LOCAL_MAX_CONCURRENT, str(DEFAULT_RERANKER_LOCAL_MAX_CONCURRENT))
+            ),
             reranker_tei_url=os.getenv(ENV_RERANKER_TEI_URL),
             reranker_tei_batch_size=int(os.getenv(ENV_RERANKER_TEI_BATCH_SIZE, str(DEFAULT_RERANKER_TEI_BATCH_SIZE))),
             reranker_tei_max_concurrent=int(
@@ -444,6 +469,9 @@ class HindsightConfig:
             consolidation_batch_size=int(
                 os.getenv(ENV_CONSOLIDATION_BATCH_SIZE, str(DEFAULT_CONSOLIDATION_BATCH_SIZE))
             ),
+            consolidation_max_tokens=int(
+                os.getenv(ENV_CONSOLIDATION_MAX_TOKENS, str(DEFAULT_CONSOLIDATION_MAX_TOKENS))
+            ),
             # Database migrations
             run_migrations_on_startup=os.getenv(ENV_RUN_MIGRATIONS_ON_STARTUP, "true").lower() == "true",
             # Database connection pool
@@ -515,7 +543,7 @@ class HindsightConfig:
 
     def log_config(self) -> None:
         """Log the current configuration (without sensitive values)."""
-        logger.info(f"Database: {self.database_url}")
+        logger.info(f"Database: {self.database_url} (schema: {self.database_schema})")
         logger.info(f"LLM: provider={self.llm_provider}, model={self.llm_model}")
         if self.retain_llm_provider or self.retain_llm_model:
             retain_provider = self.retain_llm_provider or self.llm_provider

{hindsight_api-0.4.0 → hindsight_api-0.4.1}/hindsight_api/engine/consolidation/consolidator.py

@@ -639,28 +639,27 @@ async def _find_related_observations(
     request_context: "RequestContext",
 ) -> list[dict[str, Any]]:
     """
-    Find observations related to the given query using the full recall system.
+    Find observations related to the given query using optimized recall.
 
     IMPORTANT: We do NOT filter by tags here. Consolidation needs to see ALL
     potentially related observations regardless of scope, so the LLM can
     decide on tag routing (same scope update vs cross-scope create).
 
-    This leverages:
-    - Semantic search (embedding similarity)
-    - BM25 text search (keyword matching)
-    - Entity-based retrieval (shared entities)
-    - Graph traversal (connected via entity links)
+    Uses max_tokens to naturally limit observations (no artificial count limit).
+    Includes source memories with dates for LLM context.
 
     Returns:
-        List of related observations with their tags for LLM tag routing
+        List of related observations with their tags, source memories, and dates
     """
-    # Use recall to find related observations
-    # NO tags parameter - we want ALL observations regardless of scope
-    # Use low max_tokens since we only need observations, not memories
+    # Use recall to find related observations with token budget
+    # max_tokens naturally limits how many observations are returned
+    from ...config import get_config
+
+    config = get_config()
     recall_result = await memory_engine.recall_async(
         bank_id=bank_id,
         query=query,
-        max_tokens=5000,  # Token budget for observations
+        max_tokens=config.consolidation_max_tokens,  # Token budget for observations (configurable)
         fact_type=["observation"],  # Only retrieve observations
         request_context=request_context,
         _quiet=True,  # Suppress logging
@@ -668,43 +667,82 @@ async def _find_related_observations(
     )
 
     # If no observations returned, return empty list
-    # When fact_type=["observation"], results come back in `results` field
     if not recall_result.results:
         return []
 
-    # Trust recall's relevance filtering - fetch full data for each observation
+    # Batch fetch all observations in a single query (no artificial limit)
+    observation_ids = [uuid.UUID(obs.id) for obs in recall_result.results]
+
+    rows = await conn.fetch(
+        f"""
+        SELECT id, text, proof_count, history, tags, source_memory_ids, created_at, updated_at,
+               occurred_start, occurred_end, mentioned_at
+        FROM {fq_table("memory_units")}
+        WHERE id = ANY($1) AND bank_id = $2 AND fact_type = 'observation'
+        """,
+        observation_ids,
+        bank_id,
+    )
+
+    # Build results list preserving recall order
+    id_to_row = {row["id"]: row for row in rows}
     results = []
-    for obs in recall_result.results:
-        # Fetch full observation data from DB to get history, source_memory_ids, tags
-        row = await conn.fetchrow(
-            f"""
-            SELECT id, text, proof_count, history, tags, source_memory_ids, created_at, updated_at
-            FROM {fq_table("memory_units")}
-            WHERE id = $1 AND bank_id = $2 AND fact_type = 'observation'
-            """,
-            uuid.UUID(obs.id),
-            bank_id,
-        )
 
-        if row:
-            history = row["history"]
-            if isinstance(history, str):
-                history = json.loads(history)
-            elif history is None:
-                history = []
-
-            results.append(
-                {
-                    "id": row["id"],
-                    "text": row["text"],
-                    "proof_count": row["proof_count"] or 1,
-                    "history": history,
-                    "tags": row["tags"] or [],  # Include tags for LLM tag routing
-                    "source_memory_ids": row["source_memory_ids"] or [],
-                    "similarity": 1.0,  # Retrieved via recall so assumed relevant
-                }
+    for obs in recall_result.results:
+        obs_id = uuid.UUID(obs.id)
+        if obs_id not in id_to_row:
+            continue
+
+        row = id_to_row[obs_id]
+        history = row["history"]
+        if isinstance(history, str):
+            history = json.loads(history)
+        elif history is None:
+            history = []
+
+        # Fetch source memories to include their text and dates
+        source_memory_ids = row["source_memory_ids"] or []
+        source_memories = []
+
+        if source_memory_ids:
+            source_rows = await conn.fetch(
+                f"""
+                SELECT text, occurred_start, occurred_end, mentioned_at, event_date
+                FROM {fq_table("memory_units")}
+                WHERE id = ANY($1) AND bank_id = $2
+                ORDER BY created_at ASC
+                LIMIT 5
+                """,
+                source_memory_ids[:5],  # Limit to first 5 source memories for token efficiency
+                bank_id,
             )
 
+            for src_row in source_rows:
+                source_memories.append(
+                    {
+                        "text": src_row["text"],
+                        "occurred_start": src_row["occurred_start"],
+                        "occurred_end": src_row["occurred_end"],
+                        "mentioned_at": src_row["mentioned_at"],
+                        "event_date": src_row["event_date"],
+                    }
+                )
+
+        results.append(
+            {
+                "id": row["id"],
+                "text": row["text"],
+                "proof_count": row["proof_count"] or 1,
+                "tags": row["tags"] or [],
+                "source_memories": source_memories,
+                "occurred_start": row["occurred_start"],
+                "occurred_end": row["occurred_end"],
+                "mentioned_at": row["mentioned_at"],
+                "created_at": row["created_at"],
+                "updated_at": row["updated_at"],
+            }
+        )
+
     return results
 
 
@@ -732,14 +770,43 @@ async def _consolidate_with_llm(
         - {"action": "create", "text": "...", "reason": "..."}
         - [] if fact is purely ephemeral (no durable knowledge)
     """
-    # Format observations WITH their tags (or "None" if empty)
+    # Format observations as JSON with source memories and dates
     if observations:
-        observations_text = "\n".join(
-            f'- ID: {obs["id"]}, Tags: {json.dumps(obs["tags"])}, Text: "{obs["text"]}" (proof_count: {obs["proof_count"]})'
-            for obs in observations
-        )
+        obs_list = []
+        for obs in observations:
+            obs_data = {
+                "id": str(obs["id"]),
+                "text": obs["text"],
+                "proof_count": obs["proof_count"],
+                "tags": obs["tags"],
+                "created_at": obs["created_at"].isoformat() if obs.get("created_at") else None,
+                "updated_at": obs["updated_at"].isoformat() if obs.get("updated_at") else None,
+            }
+
+            # Include temporal info if available
+            if obs.get("occurred_start"):
+                obs_data["occurred_start"] = obs["occurred_start"].isoformat()
+            if obs.get("occurred_end"):
+                obs_data["occurred_end"] = obs["occurred_end"].isoformat()
+            if obs.get("mentioned_at"):
+                obs_data["mentioned_at"] = obs["mentioned_at"].isoformat()
+
+            # Include source memories (up to 3 for brevity)
+            if obs.get("source_memories"):
+                obs_data["source_memories"] = [
+                    {
+                        "text": sm["text"],
+                        "event_date": sm["event_date"].isoformat() if sm.get("event_date") else None,
+                        "occurred_start": sm["occurred_start"].isoformat() if sm.get("occurred_start") else None,
+                    }
+                    for sm in obs["source_memories"][:3]  # Limit to 3 for token efficiency
+                ]
+
+            obs_list.append(obs_data)
+
+        observations_text = json.dumps(obs_list, indent=2)
     else:
-        observations_text = "None (this is a new topic - create if fact contains durable knowledge)"
+        observations_text = "[]"
 
     # Only include mission section if mission is set and not the default
     mission_section = ""

{hindsight_api-0.4.0 → hindsight_api-0.4.1}/hindsight_api/engine/consolidation/prompts.py

@@ -47,23 +47,31 @@ CONSOLIDATION_USER_PROMPT = """Analyze this new fact and consolidate into knowle
 {mission_section}
 NEW FACT: {fact_text}
 
-EXISTING OBSERVATIONS:
+EXISTING OBSERVATIONS (JSON array with source memories and dates):
 {observations_text}
 
-Instructions:
-1. First, extract the DURABLE KNOWLEDGE from the fact (not ephemeral state like "user is at X")
-2. Then compare with existing observations:
-   - If an observation covers the same topic: UPDATE it with the new knowledge
-   - If no observation covers the topic: CREATE a new one
+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
 
-Output JSON array of actions (ALWAYS an array, even for single action):
+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", "text": "updated durable knowledge", "reason": "..."}},
+  {{"action": "update", "learning_id": "uuid-from-observations", "text": "updated knowledge", "reason": "..."}},
   {{"action": "create", "text": "new durable knowledge", "reason": "..."}}
 ]
 
-If NO consolidation is needed (fact is purely ephemeral with no durable knowledge):
-[]
-
-If no observations exist and fact contains durable knowledge:
-[{{"action": "create", "text": "durable knowledge text", "reason": "new topic"}}]"""
+Return [] if fact contains no durable knowledge."""

{hindsight_api-0.4.0 → hindsight_api-0.4.1}/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
 
@@ -139,13 +144,23 @@ class LocalSTCrossEncoder(CrossEncoderModel):
         # after loading, which conflicts with accelerate's device_map handling.
         import torch
 
-        # Check for GPU (CUDA) or Apple Silicon (MPS)
-        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
-        else:
+        # 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,
@@ -211,12 +226,19 @@ class LocalSTCrossEncoder(CrossEncoderModel):
             )
 
         # Determine device based on hardware availability
-        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
-        else:
+        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,
@@ -873,29 +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: