Skip to content

Compress tool outputs, logs, files, and RAG chunks before they reach the LLM. 20% fewer tokens for coding agents, 60-95% fewer tokens for JSON, same answers. Library, proxy, MCP server.

Add centralized ML model configuration

chopratejas7cf10675ea85
diff --git a/headroom/cache/dynamic_detector.py b/headroom/cache/dynamic_detector.py
index d2df21ac2..039da0515 100644
--- a/headroom/cache/dynamic_detector.py
+++ b/headroom/cache/dynamic_detector.py
@@ -38,6 +38,8 @@ from dataclasses import dataclass, field
from enum import Enum
from typing import Any, Literal
+from headroom.models.config import ML_MODEL_DEFAULTS
+
# Optional imports - graceful degradation
_SPACY_AVAILABLE = False
_SENTENCE_TRANSFORMERS_AVAILABLE = False
@@ -220,13 +222,13 @@ class DetectorConfig:
min_entropy_length: int = 8
# Tier 2: NER config
- spacy_model: str = "en_core_web_sm"
+ spacy_model: str = field(default_factory=lambda: ML_MODEL_DEFAULTS.spacy)
ner_entity_types: list[str] = field(
default_factory=lambda: ["DATE", "TIME", "MONEY", "PERSON", "ORG", "GPE"]
)
# Tier 3: Semantic config
- embedding_model: str = "all-MiniLM-L6-v2"
+ embedding_model: str = field(default_factory=lambda: ML_MODEL_DEFAULTS.sentence_transformer)
semantic_threshold: float = 0.7
# General
diff --git a/headroom/cache/semantic.py b/headroom/cache/semantic.py
index 364978768..8eb7cca85 100644
--- a/headroom/cache/semantic.py
+++ b/headroom/cache/semantic.py
@@ -39,9 +39,11 @@ import hashlib
import time
from collections import OrderedDict
from collections.abc import Callable
-from dataclasses import dataclass
+from dataclasses import dataclass, field
from typing import Any
+from headroom.models.config import ML_MODEL_DEFAULTS
+
from .base import (
BaseCacheOptimizer,
CacheConfig,
@@ -90,7 +92,7 @@ class SemanticCacheConfig:
use_exact_matching: bool = True
# Embedding model (if using embeddings)
- embedding_model: str = "all-MiniLM-L6-v2"
+ embedding_model: str = field(default_factory=lambda: ML_MODEL_DEFAULTS.sentence_transformer)
class SemanticCache:
diff --git a/headroom/config.py b/headroom/config.py
index aa073fc4e..c3fbc7534 100644
--- a/headroom/config.py
+++ b/headroom/config.py
@@ -7,6 +7,8 @@ from datetime import datetime
from enum import Enum
from typing import Any, Literal
+from headroom.models.config import ML_MODEL_DEFAULTS
+
class HeadroomMode(str, Enum):
"""Operating modes for Headroom."""
@@ -275,7 +277,7 @@ class RelevanceScorerConfig:
bm25_b: float = 0.75 # Length normalization
# Embedding parameters
- embedding_model: str = "all-MiniLM-L6-v2" # Lightweight model
+ embedding_model: str = field(default_factory=lambda: ML_MODEL_DEFAULTS.sentence_transformer)
# Hybrid parameters
hybrid_alpha: float = 0.5 # BM25 weight (1-alpha = embedding weight)
diff --git a/headroom/evals/metrics.py b/headroom/evals/metrics.py
index e6bc6e248..58ccb865b 100644
--- a/headroom/evals/metrics.py
+++ b/headroom/evals/metrics.py
@@ -143,7 +143,7 @@ def compute_rouge_l(response_a: str, response_b: str) -> float:
def compute_semantic_similarity(
response_a: str,
response_b: str,
- model_name: str = "all-MiniLM-L6-v2",
+ model_name: str | None = None,
) -> float:
"""Compute semantic similarity using sentence embeddings.
@@ -152,7 +152,7 @@ def compute_semantic_similarity(
Args:
response_a: First response
response_b: Second response
- model_name: Sentence transformer model to use
+ model_name: Sentence transformer model to use. Uses config default if None.
Returns:
Cosine similarity between embeddings (0.0 to 1.0)
diff --git a/headroom/image/compressor.py b/headroom/image/compressor.py
index 9e4fa590a..8e06c4c92 100644
--- a/headroom/image/compressor.py
+++ b/headroom/image/compressor.py
@@ -67,11 +67,9 @@ class ImageCompressor:
device: Device for inference ('cuda', 'cpu', or None for auto)
"""
- DEFAULT_MODEL = "chopratejas/technique-router"
-
def __init__(
self,
- model_id: str = DEFAULT_MODEL,
+ model_id: str | None = None,
use_siglip: bool = True,
device: str | None = None,
):
diff --git a/headroom/image/trained_router.py b/headroom/image/trained_router.py
index 0ab24c0a1..c675a82e0 100644
--- a/headroom/image/trained_router.py
+++ b/headroom/image/trained_router.py
@@ -19,6 +19,8 @@ from typing import Any
import torch
from PIL import Image
+from headroom.models.config import ML_MODEL_DEFAULTS
+
class Technique(Enum):
"""Image optimization techniques."""
@@ -64,9 +66,14 @@ class TrainedRouter:
- HuggingFace Hub: headroom-ai/technique-router (for production)
"""
- # Model identifiers
- DEFAULT_HF_MODEL = "chopratejas/technique-router"
- SIGLIP_MODEL = "google/siglip-base-patch16-224"
+ # Model identifiers (from centralized config)
+ @property
+ def default_hf_model(self) -> str:
+ return ML_MODEL_DEFAULTS.technique_router
+
+ @property
+ def siglip_model(self) -> str:
+ return ML_MODEL_DEFAULTS.siglip
# Image analysis prompts for SigLIP
IMAGE_DESCRIPTIONS = {
@@ -138,7 +145,7 @@ class TrainedRouter:
if local_path.exists():
model_id = str(local_path)
else:
- model_id = self.DEFAULT_HF_MODEL
+ model_id = self.default_hf_model
# Use centralized registry for shared model instances
from headroom.models.ml_models import MLModelRegistry
@@ -153,7 +160,7 @@ class TrainedRouter:
from headroom.models.ml_models import MLModelRegistry
self._siglip_model, self._siglip_processor = MLModelRegistry.get_siglip(
- model_name=self.SIGLIP_MODEL,
+ model_name=self.siglip_model,
device=self.device,
)
diff --git a/headroom/memory/adapters/embedders.py b/headroom/memory/adapters/embedders.py
index 29385a3e8..50c6b94f5 100644
--- a/headroom/memory/adapters/embedders.py
+++ b/headroom/memory/adapters/embedders.py
@@ -18,6 +18,8 @@ from typing import TYPE_CHECKING, Any
import numpy as np
+from headroom.models.config import ML_MODEL_DEFAULTS
+
if TYPE_CHECKING:
from sentence_transformers import SentenceTransformer
@@ -81,7 +83,6 @@ class LocalEmbedder:
embeddings = await embedder.embed_batch(["Hello", "World"])
"""
- DEFAULT_MODEL = "all-MiniLM-L6-v2"
DEFAULT_DIMENSION = 384
DEFAULT_MAX_TOKENS = 256
@@ -94,14 +95,14 @@ class LocalEmbedder:
Args:
model_name: Name of the sentence-transformers model to use.
- Defaults to "all-MiniLM-L6-v2".
+ Defaults to config's sentence_transformer setting.
device: Device to run on ("cuda", "mps", "cpu", or None for auto).
If None, automatically selects the best available device.
Raises:
ImportError: If sentence-transformers is not installed.
"""
- self._model_name = model_name or self.DEFAULT_MODEL
+ self._model_name = model_name or ML_MODEL_DEFAULTS.sentence_transformer
self._requested_device = device
self._model: SentenceTransformer | None = None
self._device: str | None = None
diff --git a/headroom/memory/backends/local.py b/headroom/memory/backends/local.py
index 9346c4097..c79f0739c 100644
--- a/headroom/memory/backends/local.py
+++ b/headroom/memory/backends/local.py
@@ -15,13 +15,14 @@ from __future__ import annotations
import logging
import uuid
-from dataclasses import dataclass
+from dataclasses import dataclass, field
from pathlib import Path
from typing import TYPE_CHECKING, Any
from headroom.memory.adapters.graph_models import Entity, Relationship, Subgraph
from headroom.memory.models import Memory
from headroom.memory.ports import MemorySearchResult
+from headroom.models.config import ML_MODEL_DEFAULTS
if TYPE_CHECKING:
from headroom.memory.adapters.graph import InMemoryGraphStore
@@ -51,8 +52,10 @@ class LocalBackendConfig:
db_path: str = "memory.db"
graph_db_path: str | None = None # Derived from db_path if not specified
- embedder_model: str = "all-MiniLM-L6-v2"
- vector_dimension: int = 384
+ embedder_model: str = field(default_factory=lambda: ML_MODEL_DEFAULTS.sentence_transformer)
+ vector_dimension: int = field(
+ default_factory=lambda: ML_MODEL_DEFAULTS.sentence_transformer_dim
+ )
graph_persist: bool = True # Use SQLiteGraphStore (bounded, persistent)
graph_cache_size_kb: int = 8192 # 8MB default
cache_enabled: bool = True
diff --git a/headroom/memory/config.py b/headroom/memory/config.py
index 49a15823c..326c47122 100644
--- a/headroom/memory/config.py
+++ b/headroom/memory/config.py
@@ -15,6 +15,8 @@ from dataclasses import dataclass, field
from enum import Enum
from pathlib import Path
+from headroom.models.config import ML_MODEL_DEFAULTS
+
class StoreBackend(Enum):
"""Supported memory store backends."""
@@ -112,7 +114,7 @@ class MemoryConfig:
# Embedder
embedder_backend: EmbedderBackend = EmbedderBackend.LOCAL
- embedder_model: str = "all-MiniLM-L6-v2"
+ embedder_model: str = field(default_factory=lambda: ML_MODEL_DEFAULTS.sentence_transformer)
openai_api_key: str | None = None
ollama_base_url: str = "http://localhost:11434"
diff --git a/headroom/models/config.py b/headroom/models/config.py
new file mode 100644
index 000000000..f6fbb6332
--- /dev/null
+++ b/headroom/models/config.py
@@ -0,0 +1,167 @@
+"""Central configuration for all ML models used in Headroom.
+
+This is the SINGLE SOURCE OF TRUTH for model defaults. Change values here
+to switch model variants across the entire codebase.
+
+Usage:
+ from headroom.models.config import ML_MODEL_DEFAULTS
+
+ # Get default model name
+ model = ML_MODEL_DEFAULTS.sentence_transformer
+
+ # Or use environment variables to override at runtime:
+ # HEADROOM_SENTENCE_TRANSFORMER=intfloat/e5-small-v2
+ # HEADROOM_SIGLIP=google/siglip-base-patch16-224
+ # HEADROOM_LLMLINGUA=microsoft/llmlingua-2-bert-base-multilingual-cased-meetingbank
+"""
+
+from __future__ import annotations
+
+import os
+from dataclasses import dataclass, field
+
+
+@dataclass
+class MLModelConfig:
+ """Central configuration for all ML model defaults.
+
+ All model names are defined here. Components throughout Headroom
+ import these defaults, so changing a value here changes it everywhere.
+
+ Environment variables can override any default:
+ - HEADROOM_SENTENCE_TRANSFORMER
+ - HEADROOM_SIGLIP
+ - HEADROOM_LLMLINGUA
+ - HEADROOM_SPACY
+ - HEADROOM_TECHNIQUE_ROUTER
+
+ Attributes:
+ sentence_transformer: Model for text embeddings (semantic similarity, memory).
+ Default: all-MiniLM-L6-v2 (22M params, 384 dim, ~90MB)
+ Alternative: intfloat/e5-small-v2 (33M params, better accuracy)
+
+ sentence_transformer_dim: Embedding dimension for the sentence transformer.
+ Must match the model's output dimension.
+
+ siglip: Model for image embeddings and analysis.
+ Default: google/siglip-base-patch16-224 (~400MB)
+ Alternative: google/siglip-so400m-patch14-384 (larger, more accurate)
+
+ llmlingua: Model for ML-based prompt compression.
+ Default: microsoft/llmlingua-2-bert-base-multilingual-cased-meetingbank (~350MB)
+ Alternative: microsoft/llmlingua-2-xlm-roberta-large-meetingbank (~1GB, slightly more accurate)
+
+ spacy: Model for named entity recognition.
+ Default: en_core_web_sm (~40MB)
+ Alternative: en_core_web_md (~120MB, more accurate)
+
+ technique_router: Model for image optimization routing.
+ Default: chopratejas/technique-router (~100MB)
+ """
+
+ # Text Embeddings (SentenceTransformer)
+ sentence_transformer: str = field(
+ default_factory=lambda: os.environ.get("HEADROOM_SENTENCE_TRANSFORMER", "all-MiniLM-L6-v2")
+ )
+ sentence_transformer_dim: int = 384
+
+ # Image Embeddings (SIGLIP)
+ siglip: str = field(
+ default_factory=lambda: os.environ.get("HEADROOM_SIGLIP", "google/siglip-base-patch16-224")
+ )
+
+ # Prompt Compression (LLMLingua-2)
+ llmlingua: str = field(
+ default_factory=lambda: os.environ.get(
+ "HEADROOM_LLMLINGUA", "microsoft/llmlingua-2-bert-base-multilingual-cased-meetingbank"
+ )
+ )
+
+ # Named Entity Recognition (spaCy)
+ spacy: str = field(default_factory=lambda: os.environ.get("HEADROOM_SPACY", "en_core_web_sm"))
+
+ # Image Technique Router
+ technique_router: str = field(
+ default_factory=lambda: os.environ.get(
+ "HEADROOM_TECHNIQUE_ROUTER", "chopratejas/technique-router"
+ )
+ )
+
+ # Memory estimates in MB (for monitoring)
+ _memory_estimates: dict[str, int] = field(
+ default_factory=lambda: {
+ # Sentence Transformers
+ "all-MiniLM-L6-v2": 90,
+ "all-mpnet-base-v2": 420,
+ "intfloat/e5-small-v2": 130,
+ "intfloat/e5-base-v2": 440,
+ # SIGLIP
+ "google/siglip-base-patch16-224": 400,
+ "google/siglip-so400m-patch14-384": 900,
+ "google/siglip-large-patch16-384": 1200,
+ # LLMLingua
+ "microsoft/llmlingua-2-xlm-roberta-large-meetingbank": 1000,
+ "microsoft/llmlingua-2-bert-base-multilingual-cased-meetingbank": 350,
+ # spaCy
+ "en_core_web_sm": 40,
+ "en_core_web_md": 120,
+ "en_core_web_lg": 560,
+ # Technique Router
+ "chopratejas/technique-router": 100,
+ }
+ )
+
+ def get_memory_estimate(self, model_name: str) -> int:
+ """Get estimated memory usage for a model in MB.
+
+ Args:
+ model_name: The model identifier.
+
+ Returns:
+ Estimated memory in MB, or 100 if unknown.
+ """
+ return self._memory_estimates.get(model_name, 100)
+
+ def total_memory_estimate(self) -> int:
+ """Get total estimated memory if all configured models are loaded.
+
+ Returns:
+ Total estimated memory in MB.
+ """
+ return (
+ self.get_memory_estimate(self.sentence_transformer)
+ + self.get_memory_estimate(self.siglip)
+ + self.get_memory_estimate(self.llmlingua)
+ + self.get_memory_estimate(self.spacy)
+ + self.get_memory_estimate(self.technique_router)
+ )
+
+
+# Singleton instance - import this to get defaults
+ML_MODEL_DEFAULTS = MLModelConfig()
+
+
+# Convenience accessors for common use cases
+def get_default_embedding_model() -> str:
+ """Get the default sentence transformer model name."""
+ return ML_MODEL_DEFAULTS.sentence_transformer
+
+
+def get_default_embedding_dim() -> int:
+ """Get the default embedding dimension."""
+ return ML_MODEL_DEFAULTS.sentence_transformer_dim
+
+
+def get_default_spacy_model() -> str:
+ """Get the default spaCy model name."""
+ return ML_MODEL_DEFAULTS.spacy
+
+
+def get_default_siglip_model() -> str:
+ """Get the default SIGLIP model name."""
+ return ML_MODEL_DEFAULTS.siglip
+
+
+def get_default_llmlingua_model() -> str:
+ """Get the default LLMLingua model name."""
+ return ML_MODEL_DEFAULTS.llmlingua
diff --git a/headroom/models/ml_models.py b/headroom/models/ml_models.py
index 056f9399c..50cc3d17c 100644
--- a/headroom/models/ml_models.py
+++ b/headroom/models/ml_models.py
@@ -6,10 +6,13 @@ to avoid loading the same model multiple times across different components.
This is different from registry.py which stores LLM metadata. This module
manages actual loaded model instances that consume memory.
+Model defaults are configured in headroom/models/config.py - change them there
+to switch model variants across the entire codebase.
+
Usage:
from headroom.models.ml_models import MLModelRegistry
- # Get shared sentence transformer (loads on first access)
+ # Get shared sentence transformer (loads on first access, uses config default)
model = MLModelRegistry.get_sentence_transformer()
embeddings = model.encode(["hello", "world"])
@@ -27,23 +30,13 @@ import logging
from threading import RLock
from typing import TYPE_CHECKING, Any
+from .config import ML_MODEL_DEFAULTS
+
if TYPE_CHECKING:
pass
logger = logging.getLogger(__name__)
-# Model size estimates in MB (approximate)
-MODEL_SIZES_MB = {
- "sentence_transformer:all-MiniLM-L6-v2": 90,
- "sentence_transformer:all-mpnet-base-v2": 420,
- "siglip:google/siglip-base-patch16-224": 400,
- "siglip:google/siglip-large-patch16-384": 1200,
- "llmlingua:microsoft/llmlingua-2-xlm-roberta-large-meetingbank": 1000,
- "spacy:en_core_web_sm": 40,
- "spacy:en_core_web_md": 120,
- "technique_router:chopratejas/technique-router": 100,
-}
-
class MLModelRegistry:
"""Singleton registry for shared ML model instances.
@@ -90,18 +83,21 @@ class MLModelRegistry:
@classmethod
def get_sentence_transformer(
cls,
- model_name: str = "all-MiniLM-L6-v2",
+ model_name: str | None = None,
device: str | None = None,
) -> Any:
"""Get a shared SentenceTransformer instance.
Args:
- model_name: Model name (default: all-MiniLM-L6-v2).
+ model_name: Model name. If None, uses ML_MODEL_DEFAULTS.sentence_transformer.
device: Device to use (cuda, mps, cpu). Auto-detected if None.
Returns:
SentenceTransformer model instance.
"""
+ if model_name is None:
+ model_name = ML_MODEL_DEFAULTS.sentence_transformer
+
instance = cls.get()
key = f"sentence_transformer:{model_name}"
@@ -126,18 +122,21 @@ class MLModelRegistry:
@classmethod
def get_siglip(
cls,
- model_name: str = "google/siglip-base-patch16-224",
+ model_name: str | None = None,
device: str | None = None,
) -> tuple[Any, Any]:
"""Get shared SIGLIP model and processor.
Args:
- model_name: Model name (default: google/siglip-base-patch16-224).
+ model_name: Model name. If None, uses ML_MODEL_DEFAULTS.siglip.
device: Device to use. Auto-detected if None.
Returns:
Tuple of (model, processor).
"""
+ if model_name is None:
+ model_name = ML_MODEL_DEFAULTS.siglip
+
instance = cls.get()
key = f"siglip:{model_name}"
@@ -170,15 +169,18 @@ class MLModelRegistry:
# =========================================================================
@classmethod
- def get_spacy(cls, model_name: str = "en_core_web_sm") -> Any:
+ def get_spacy(cls, model_name: str | None = None) -> Any:
"""Get a shared spaCy model.
Args:
- model_name: Model name (default: en_core_web_sm).
+ model_name: Model name. If None, uses ML_MODEL_DEFAULTS.spacy.
Returns:
spaCy Language model.
"""
+ if model_name is None:
+ model_name = ML_MODEL_DEFAULTS.spacy
+
instance = cls.get()
key = f"spacy:{model_name}"
@@ -222,7 +224,7 @@ class MLModelRegistry:
if local_path.exists():
model_path = str(local_path)
else:
- model_path = "chopratejas/technique-router"
+ model_path = ML_MODEL_DEFAULTS.technique_router
key = f"technique_router:{model_path}"
@@ -274,7 +276,7 @@ class MLModelRegistry:
device = cls._detect_device()
if model_name is None:
- model_name = "microsoft/llmlingua-2-xlm-roberta-large-meetingbank"
+ model_name = ML_MODEL_DEFAULTS.llmlingua
return _get_llmlingua_compressor(model_name=model_name, device=device)
@@ -317,7 +319,9 @@ class MLModelRegistry:
total = 0.0
with instance._model_lock:
for key in instance._models:
- total += MODEL_SIZES_MB.get(key, 100) # Default 100MB if unknown
+ # Extract model name from key (format: "type:model_name")
+ model_name = key.split(":", 1)[1] if ":" in key else key
+ total += ML_MODEL_DEFAULTS.get_memory_estimate(model_name)
return total
@classmethod
@@ -329,7 +333,9 @@ class MLModelRegistry:
with instance._model_lock:
for key in instance._models:
- size_mb = MODEL_SIZES_MB.get(key, 100)
+ # Extract model name from key (format: "type:model_name")
+ model_name = key.split(":", 1)[1] if ":" in key else key
+ size_mb = ML_MODEL_DEFAULTS.get_memory_estimate(model_name)
loaded_models.append({"key": key, "size_mb": size_mb})
total_estimated_mb += size_mb
@@ -341,7 +347,7 @@ class MLModelRegistry:
# Convenience functions for direct access
def get_sentence_transformer(
- model_name: str = "all-MiniLM-L6-v2",
+ model_name: str | None = None,
device: str | None = None,
) -> Any:
"""Get a shared SentenceTransformer instance."""
@@ -349,13 +355,13 @@ def get_sentence_transformer(
def get_siglip(
- model_name: str = "google/siglip-base-patch16-224",
+ model_name: str | None = None,
device: str | None = None,
) -> tuple[Any, Any]:
"""Get shared SIGLIP model and processor."""
return MLModelRegistry.get_siglip(model_name, device)
-def get_spacy(model_name: str = "en_core_web_sm") -> Any:
+def get_spacy(model_name: str | None = None) -> Any:
"""Get a shared spaCy model."""
return MLModelRegistry.get_spacy(model_name)
diff --git a/headroom/prediction/feature_extractor.py b/headroom/prediction/feature_extractor.py
index fb019a14d..a0222524e 100644
--- a/headroom/prediction/feature_extractor.py
+++ b/headroom/prediction/feature_extractor.py
@@ -39,6 +39,8 @@ from dataclasses import dataclass, field
from enum import Enum
from typing import TYPE_CHECKING, Any, ClassVar
+from headroom.models.config import ML_MODEL_DEFAULTS
+
if TYPE_CHECKING:
from sentence_transformers import SentenceTransformer
@@ -1744,7 +1746,7 @@ class SemanticExtractor(BaseFeatureExtractor):
# Use centralized registry for shared model instances
from headroom.models.ml_models import MLModelRegistry
- self._nlp = MLModelRegistry.get_spacy("en_core_web_sm")
+ self._nlp = MLModelRegistry.get_spacy()
assert self._nlp is not None
doc = self._nlp(text)
@@ -1952,18 +1954,18 @@ class EmbeddingExtractor(BaseFeatureExtractor):
def __init__(
self,
- model_name: str = "all-MiniLM-L6-v2",
+ model_name: str | None = None,
device: str | None = None,
cluster_centers: dict[str, list[float]] | None = None,
):
"""Initialize embedding extractor.
Args:
- model_name: Sentence transformer model name.
+ model_name: Sentence transformer model name. Uses config default if None.
device: Device for model ('cpu', 'cuda', 'mps', or None for auto).
cluster_centers: Pre-computed cluster centers for similarity.
"""
- self.model_name = model_name
+ self.model_name = model_name or ML_MODEL_DEFAULTS.sentence_transformer
self.device = device
self.cluster_centers = cluster_centers or self.DEFAULT_CLUSTERS
self._model: SentenceTransformer | None = None
@@ -2260,7 +2262,7 @@ class PromptFeatureExtractor:
tokenizer: Any | None = None,
use_embeddings: bool = True,
use_ner: bool = False,
- embedding_model: str = "all-MiniLM-L6-v2",
+ embedding_model: str | None = None,
cluster_centers: dict[str, list[float]] | None = None,
):
"""Initialize the feature extractor.
@@ -2272,7 +2274,7 @@ class PromptFeatureExtractor:
Requires sentence-transformers.
use_ner: Whether to use NER for entity extraction.
Requires spaCy.
- embedding_model: Sentence transformer model name.
+ embedding_model: Sentence transformer model name. Uses config default if None.
cluster_centers: Pre-computed cluster centers for similarity.
"""
self.text_extractor = TextStatisticsExtractor(tokenizer=tokenizer)
diff --git a/headroom/relevance/embedding.py b/headroom/relevance/embedding.py
index 22953ec5c..4c85c3a53 100644
--- a/headroom/relevance/embedding.py
+++ b/headroom/relevance/embedding.py
@@ -92,14 +92,14 @@ class EmbeddingScorer(RelevanceScorer):
def __init__(
self,
- model_name: str = "all-MiniLM-L6-v2",
+ model_name: str | None = None,
device: str | None = None,
cache_model: bool = True, # Kept for API compatibility, always uses registry now
):
"""Initialize embedding scorer.
Args:
- model_name: Sentence transformer model name.
+ model_name: Sentence transformer model name. If None, uses config default.
Recommended models:
- "all-MiniLM-L6-v2": Fast, good quality (default)
- "all-mpnet-base-v2": Best quality, slower
@@ -107,7 +107,9 @@ class EmbeddingScorer(RelevanceScorer):
device: Device to use ('cpu', 'cuda', 'mps', or None for auto).
cache_model: Deprecated, models are always cached via MLModelRegistry.
"""
- self.model_name = model_name
+ from headroom.models.config import ML_MODEL_DEFAULTS
+
+ self.model_name = model_name or ML_MODEL_DEFAULTS.sentence_transformer
self.device = device
self.cache_model = cache_model
self._available: bool | None = None
diff --git a/headroom/transforms/llmlingua_compressor.py b/headroom/transforms/llmlingua_compressor.py
index c760399e8..26b3861f4 100644
--- a/headroom/transforms/llmlingua_compressor.py
+++ b/headroom/transforms/llmlingua_compressor.py
@@ -32,6 +32,7 @@ from dataclasses import dataclass, field
from typing import Any
from ..config import TransformResult
+from ..models.config import ML_MODEL_DEFAULTS
from ..tokenizer import Tokenizer
from .base import Transform
@@ -199,7 +200,7 @@ class LLMLinguaConfig:
"""
# Model configuration
- model_name: str = "microsoft/llmlingua-2-xlm-roberta-large-meetingbank"
+ model_name: str = field(default_factory=lambda: ML_MODEL_DEFAULTS.llmlingua)
device: str = "auto"
# Compression parameters
diff --git a/tests/test_memory/test_factory.py b/tests/test_memory/test_factory.py
index 73fcbea3e..4da9df11a 100644
--- a/tests/test_memory/test_factory.py
+++ b/tests/test_memory/test_factory.py
@@ -388,9 +388,11 @@ class TestCreateMemorySystem:
from headroom.memory.adapters.fts5 import FTS5TextIndex
from headroom.memory.adapters.hnsw import HNSWVectorIndex
from headroom.memory.adapters.sqlite import SQLiteMemoryStore
+ from headroom.memory.adapters.sqlite_vector import SQLiteVectorIndex
assert isinstance(store, SQLiteMemoryStore)
- assert isinstance(vector, HNSWVectorIndex)
+ # Factory auto-selects best available: SQLiteVectorIndex (preferred) or HNSW (fallback)
+ assert isinstance(vector, (SQLiteVectorIndex, HNSWVectorIndex))
assert isinstance(text, FTS5TextIndex)
assert isinstance(embedder, LocalEmbedder)
assert isinstance(cache, LRUMemoryCache)
@@ -416,9 +418,11 @@ class TestCreateMemorySystem:
from headroom.memory.adapters.fts5 import FTS5TextIndex
from headroom.memory.adapters.hnsw import HNSWVectorIndex
from headroom.memory.adapters.sqlite import SQLiteMemoryStore
+ from headroom.memory.adapters.sqlite_vector import SQLiteVectorIndex
assert isinstance(store, SQLiteMemoryStore)
- assert isinstance(vector, HNSWVectorIndex)
+ # Factory auto-selects best available: SQLiteVectorIndex (preferred) or HNSW (fallback)
+ assert isinstance(vector, (SQLiteVectorIndex, HNSWVectorIndex))
assert isinstance(text, FTS5TextIndex)
assert isinstance(embedder, LocalEmbedder)
assert isinstance(cache, LRUMemoryCache)
@@ -457,6 +461,7 @@ class TestCreateMemorySystem:
with tempfile.TemporaryDirectory() as tmpdir:
config = MemoryConfig(
db_path=Path(tmpdir) / "test.db",
+ vector_backend=VectorBackend.HNSW, # Explicitly use HNSW to test params
vector_dimension=512,
hnsw_ef_construction=300,
hnsw_m=24,
diff --git a/tests/test_memory_integration.py b/tests/test_memory_integration.py
index 19d958b53..a80f58df5 100644
--- a/tests/test_memory_integration.py
+++ b/tests/test_memory_integration.py
@@ -210,16 +210,14 @@ class TestMemoryIntegration:
assert john_entity is not None
assert john_entity.entity_type == "person"
- # Verify relationships were added
- # Get all relationships by counting via primary storage
- all_rels = [r for r in graph._relationships.values() if r.user_id == user_id]
- assert len(all_rels) >= 3, f"Expected at least 3 relationships, got {len(all_rels)}"
+ # Verify relationships were added by querying via public API
+ from headroom.memory.adapters.graph_models import RelationshipDirection
- # Verify specific relationship exists (John -> Netflix)
+ # Verify John has outgoing relationships
john_id = john_entity.id
- john_outgoing_rel_ids = graph._relationships_by_source.get(john_id, set())
- assert len(john_outgoing_rel_ids) >= 3, (
- "Expected John to have at least 3 outgoing relationships"
+ john_rels = await graph.get_relationships(john_id, RelationshipDirection.OUTGOING)
+ assert len(john_rels) >= 3, (
+ f"Expected John to have at least 3 outgoing relationships, got {len(john_rels)}"
)
await backend.close()