Centralised embedding dimensions to 1 variable

apps/knowledge/admin.py

@@ -28,7 +28,7 @@ class RoleRagDocumentAdmin(admin.ModelAdmin):
             fields.remove('embedding')
         return fields
 
-    @admin.display(description=_("Embedding Preview (1536d)"))
+    @admin.display(description=_("Embedding Preview"))
     def display_embedding(self, obj):
         if obj.embedding is not None:
             preview = list(obj.embedding[:5])

apps/knowledge/migrations/0001_initial.py

@@ -48,7 +48,7 @@ class Migration(migrations.Migration):
                 ('updated_at', models.DateTimeField(auto_now=True, verbose_name='Updated At')),
                 ('content', models.TextField()),
                 ('content_hash', models.CharField(db_index=True, max_length=64)),
-                ('embedding', pgvector.django.vector.VectorField(blank=True, dimensions=1536, null=True)),
+                ('embedding', pgvector.django.vector.VectorField(blank=True, dimensions=getattr(settings, 'EMBEDDING_DIMENSIONS', 768), null=True)),
                 ('metadata', models.JSONField(blank=True, default=dict)),
                 ('chunk_index', models.IntegerField(default=0)),
                 ('is_active', models.BooleanField(default=True)),

apps/knowledge/models.py

@@ -1,5 +1,6 @@
 import os
 
+from django.conf import settings
 from django.db import transaction
 from django.db.models import CASCADE, BooleanField, CharField, FileField, ForeignKey, IntegerField, JSONField, Model, TextField
 from django.db.models.signals import post_delete, post_save
@@ -46,7 +47,7 @@ class RoleRagDocument(IdentifierMixin, TimeStampMixin, Model):
     content = TextField()
     content_hash = CharField(max_length=64, db_index=True)
     
-    embedding = VectorField(dimensions=1536, null=True, blank=True)
+    embedding = VectorField(dimensions=settings.EMBEDDING_DIMENSIONS, null=True, blank=True)
     
     metadata = JSONField(default=dict, blank=True)
     chunk_index = IntegerField(default=0)

apps/knowledge/tasks.py

@@ -50,6 +50,8 @@ def ingest_training_file_task(self, file_uuid):
     file_obj.status = 'ingesting'
     file_obj.save()
 
+    target_dimensions = RoleRagDocument._meta.get_field('embedding').dimensions
+
     try:
         raw_text = _extract_text_from_training_file(file_obj)
         if not raw_text:
@@ -65,7 +67,11 @@ def ingest_training_file_task(self, file_uuid):
             for text_segment in _get_text_chunks(raw_text):
                 response = client.post(
                     settings.INFERENCE_SEMANTIC_CHUNK_ENDPOINT,
-                    json={"text": text_segment, "threshold": 95}
+                    json={
+                        "text": text_segment,
+                        "threshold": 95,
+                        "target_dimensions": target_dimensions,
+                    },
                 )
                 response.raise_for_status()
                 result = response.json()

apps/onboarding/mcp.py

@@ -68,10 +68,14 @@ class MCPRouter:
 
     async def _get_embedding(self, text):
         logger.info('MCP embedding request started')
+        target_dimensions = RoleRagDocument._meta.get_field('embedding').dimensions
         async with httpx.AsyncClient() as client:
             response = await client.post(
                 settings.INFERENCE_EMBEDDINGS_ENDPOINT,
-                json={'input': text},
+                json={
+                    'input': text,
+                    'target_dimensions': target_dimensions,
+                },
             )
             response.raise_for_status()
             embedding = response.json()['data'][0]['embedding']

config/settings.py

@@ -32,6 +32,7 @@ INFERENCE_SEMANTIC_CHUNK_ENDPOINT = f"{INFERENCE_URL}/v1/semantic-chunk"
 INFERENCE_EMBEDDINGS_ENDPOINT = f"{INFERENCE_URL}/v1/embeddings"
 INFERENCE_CHAT_COMPLETIONS_ENDPOINT = f"{INFERENCE_URL}/v1/chat/completions"
 INFERENCE_INGEST_TIMEOUT = float(os.getenv('INFERENCE_INGEST_TIMEOUT', '600'))
+EMBEDDING_DIMENSIONS = int(os.getenv('EMBEDDING_DIMENSIONS', '768'))
 
 STATIC_URL  = os.getenv('DJANGO_STATIC_URL',  '/static/')
 MEDIA_URL   = os.getenv('DJANGO_MEDIA_URL',  '/media/')

gpu_server.py

@@ -20,7 +20,6 @@ logger = logging.getLogger("gpu-node")
 
 EMBED_MODEL_NAME = "nomic-ai/nomic-embed-text-v1.5"
 LLM_MODEL_PATH = os.getenv("LLM_MODEL_PATH", "/app/models/Meta-Llama-3.1-8B-Instruct-Q4_K_M.gguf")
-TARGET_DIMENSIONS = 1536
 
 state: Dict[str, Any] = {}
 
@@ -79,13 +78,29 @@ async def health():
     }
 
 
-def pad_and_normalize(embeddings: torch.Tensor) -> torch.Tensor:
-    """Standardizes vector dimensions to 1536 for pgvector compatibility."""
+def _resolve_target_dimensions(payload: Dict[str, Any]) -> int:
+    raw_target = payload.get("target_dimensions")
+    if raw_target in (None, ""):
+        raise HTTPException(status_code=400, detail="'target_dimensions' is required")
+
+    try:
+        target = int(raw_target)
+    except (TypeError, ValueError) as exc:
+        raise HTTPException(status_code=400, detail="'target_dimensions' must be an integer") from exc
+
+    if target <= 0:
+        raise HTTPException(status_code=400, detail="'target_dimensions' must be > 0")
+
+    return target
+
+
+def pad_and_normalize(embeddings: torch.Tensor, target_dimensions: int) -> torch.Tensor:
+    """Dimension standardization plus L2 normalization."""
     curr_dim = embeddings.shape[1]
-    if curr_dim < TARGET_DIMENSIONS:
-        embeddings = F.pad(embeddings, (0, TARGET_DIMENSIONS - curr_dim), "constant", 0)
-    elif curr_dim > TARGET_DIMENSIONS:
-        embeddings = embeddings[:, :TARGET_DIMENSIONS]
+    if curr_dim < target_dimensions:
+        embeddings = F.pad(embeddings, (0, target_dimensions - curr_dim), "constant", 0)
+    elif curr_dim > target_dimensions:
+        embeddings = embeddings[:, :target_dimensions]
     return F.normalize(embeddings, p=2, dim=1)
 
 
@@ -94,6 +109,7 @@ async def embeddings(request: Request):
     """Generates text embeddings compatible with OpenAI API format."""
     data = await request.json()
     input_data = data.get("input", "")
+    target_dimensions = _resolve_target_dimensions(data)
 
     if isinstance(input_data, str):
         inputs = [input_data]
@@ -121,7 +137,7 @@ async def embeddings(request: Request):
 
     with torch.no_grad():
         vectors = model.encode(prefixed_inputs, convert_to_tensor=True)
-        vectors = pad_and_normalize(vectors)
+        vectors = pad_and_normalize(vectors, target_dimensions=target_dimensions)
 
     vector_list = vectors.cpu().tolist()
 
@@ -148,6 +164,7 @@ async def semantic_chunk(request: Request):
     data = await request.json()
     raw_text = data.get("text", "")
     threshold_percentile = data.get("threshold", 95)
+    target_dimensions = _resolve_target_dimensions(data)
 
     if not raw_text:
         return {"chunks": [], "embeddings": []}
@@ -162,9 +179,11 @@ async def semantic_chunk(request: Request):
     # Split by sentences
     sentences = [s.strip() for s in raw_text.replace('\n', ' ').split('. ') if s.strip()]
     if len(sentences) < 2:
+        single = model.encode([f"search_document: {raw_text}"], convert_to_tensor=True)
+        single = pad_and_normalize(single, target_dimensions=target_dimensions)
         return {
             "chunks": [raw_text],
-            "embeddings": model.encode([f"search_document: {raw_text}"]).tolist()
+            "embeddings": single.cpu().tolist(),
         }
 
     # Generate sentence embeddings to find breakpoints via cosine distance
@@ -189,7 +208,7 @@ async def semantic_chunk(request: Request):
             [f"search_document: {c}" for c in chunks],
             convert_to_tensor=True
         )
-        final_embeddings = pad_and_normalize(final_embeddings)
+        final_embeddings = pad_and_normalize(final_embeddings, target_dimensions=target_dimensions)
 
     return {
         "chunks": chunks,

report/report.tex

@@ -461,7 +461,7 @@ embeddings. This avoids naive fixed-size splits that can break context
 mid-concept.
 
 \underline{Vector storage and retrieval with pgvector}\\
-Returned chunk embeddings are stored in RoleRagDocument.embedding (1536
+Returned chunk embeddings are stored in RoleRagDocument.embedding (768
 dimensions) in PostgreSQL using pgvector, linked relationally to role
 and source file metadata. Retrieval is performed in SQL using
 cosine-distance ranking and top-k selection, allowing role filtering and