fix: minor embed compat fixes

application/scripts/reembed.py

@@ -202,7 +202,7 @@ def reembed_faiss(source_id: str, batch_size: int, dry_run: bool) -> Tuple[int,
         ``(chunks_seen, chunks_written)``.
     """
     store = VectorCreator.create_vectorstore(
-        "faiss", source_id=source_id, embeddings_key="embeddings"
+        "faiss", source_id=source_id, embeddings_key=settings.EMBEDDINGS_KEY
     )
     chunks: List[Dict[str, Any]] = store.get_chunks() or []
     if dry_run or not chunks:
@@ -217,7 +217,10 @@ def reembed_faiss(source_id: str, batch_size: int, dry_run: bool) -> Tuple[int,
     # index and touches nothing on disk. The existing index is only replaced by
     # ``save_local`` below, so a failure while embedding leaves it intact.
     rebuilt = VectorCreator.create_vectorstore(
-        "faiss", source_id=source_id, embeddings_key="embeddings", docs_init=docs
+        "faiss",
+        source_id=source_id,
+        embeddings_key=settings.EMBEDDINGS_KEY,
+        docs_init=docs,
     )
     rebuilt.save_local()
     return len(chunks), len(docs)

application/vectorstore/base.py

@@ -13,6 +13,11 @@ from application.vectorstore.model_registry import (
 )
 
 
+def _embeddings_name_is_explicit() -> bool:
+    """True when ``EMBEDDINGS_NAME`` was configured rather than defaulted."""
+    return "EMBEDDINGS_NAME" in getattr(settings, "model_fields_set", set())
+
+
 class RemoteEmbeddings:
     """
     Wrapper for remote embeddings API (OpenAI-compatible).
@@ -47,10 +52,18 @@ class RemoteEmbeddings:
         model contributes its own context window, so a request that the server
         would reject -- or silently truncate -- is clipped here instead of
         being sent and paid for.
+
+        That fallback needs the name to mean something. For a remote server it
+        is only a label forwarded as the ``model`` field, so the settings
+        default must not lend the server mpnet's 384-token window: a name
+        nobody chose describes nothing, and clipping on it would silently
+        discard most of every chunk.
         """
         configured = settings.EMBEDDINGS_MAX_INPUT_TOKENS
         if configured and configured > 0:
             return configured
+        if not _embeddings_name_is_explicit():
+            return None
         model_limit = max_input_tokens_for(self.model_name)
         return model_limit if model_limit and model_limit > 0 else None
 

application/vectorstore/embeddings_local.py

@@ -94,6 +94,27 @@ def _spec_for(model_name: str) -> EmbeddingModel:
     )
 
 
+def _pad_to_longest_in_batch(model: Any) -> None:
+    """Undo a fixed padding width baked into a model's ``tokenizer.json``."""
+    # FastEmbed enables padding only when the tokenizer declares none, so a
+    # fixed ``length`` survives loading. Shorter inputs are then padded to that
+    # width while longer ones keep their own, the batch is ragged, and the ONNX
+    # tensor build fails. mpnet ships ``length: 128``; granite does not.
+    # Mean pooling masks pad tokens, so the vectors are unaffected.
+    tokenizer = getattr(getattr(model, "model", None), "tokenizer", None)
+    padding = getattr(tokenizer, "padding", None)
+    if not isinstance(padding, dict) or padding.get("length") is None:
+        return
+    tokenizer.enable_padding(
+        direction=padding.get("direction", "right"),
+        pad_id=padding.get("pad_id", 0),
+        pad_type_id=padding.get("pad_type_id", 0),
+        pad_token=padding.get("pad_token", "<pad>"),
+        length=None,
+        pad_to_multiple_of=padding.get("pad_to_multiple_of"),
+    )
+
+
 class EmbeddingsWrapper:
     """Runs an embedding model locally through FastEmbed.
 
@@ -131,6 +152,7 @@ class EmbeddingsWrapper:
                 f"{exc}. Known models: {', '.join(known_names())}."
             ) from exc
 
+        _pad_to_longest_in_batch(self.model)
         self.dimension = self.spec.dimension or self._probe_dimension()
         logger.info("Embeddings model ready (dimension=%d)", self.dimension)