fix(server): apply default_instruction in the remaining flash embedders

packages/sie_server/src/sie_server/adapters/_utils.py

@@ -163,6 +163,24 @@ def resolve_embedding_options(
     )
 
 
+def resolve_query_instruction(
+    instruction: str | None,
+    options: dict[str, Any] | None,
+    *,
+    is_query: bool,
+) -> str | None:
+    """Return the instruction to format into the texts for ``extract_texts``.
+
+    A request instruction always wins, including an explicit ``""``. When the
+    request gives none, queries fall back to the ``default_instruction``
+    runtime option (the profile default, or a request override merged into
+    ``options``). Documents never receive the default.
+    """
+    if instruction is not None or not is_query:
+        return instruction
+    return (options or {}).get("default_instruction")
+
+
 # ---------------------------------------------------------------------------
 # Score-pair grouping (shared by ColBERT-family adapters)
 # ---------------------------------------------------------------------------

packages/sie_server/src/sie_server/adapters/bert_flash/__init__.py

@@ -30,6 +30,7 @@
 from sie_server.adapters._utils import (
     extract_texts,
     resolve_embedding_options,
+    resolve_query_instruction,
     validate_output_types,
 )
 from sie_server.core.inference_output import EncodeOutput
@@ -212,7 +213,8 @@ def encode(
         Args:
             items: List of items to encode.
             output_types: Which outputs to compute (only "dense" supported).
-            instruction: Optional instruction prefix.
+            instruction: Optional instruction prefix. For queries, ``None``
+                falls back to the ``default_instruction`` runtime option.
             is_query: Whether items are queries (affects template selection).
             prepared_items: Not used by this adapter.
 
@@ -236,7 +238,7 @@ def encode(
 
         texts = extract_texts(
             items,
-            instruction,
+            resolve_query_instruction(instruction, options, is_query=is_query),
             is_query=is_query,
             query_template=query_template,
             doc_template=doc_template,

packages/sie_server/src/sie_server/adapters/gte_sparse_flash/__init__.py

@@ -10,7 +10,12 @@
 from sie_server.adapters._flash_base import FlashBaseAdapter
 from sie_server.adapters._spec import AdapterSpec
 from sie_server.adapters._types import ERR_NOT_LOADED, ERR_REQUIRES_TEXT, ComputePrecision
-from sie_server.adapters._utils import apply_rotary_pos_emb, extract_texts, validate_output_types
+from sie_server.adapters._utils import (
+    apply_rotary_pos_emb,
+    extract_texts,
+    resolve_query_instruction,
+    validate_output_types,
+)
 from sie_server.adapters.peft_lora_mixin import PEFTLoRAMixin
 from sie_server.core.inference_output import EncodeOutput, SparseVector
 from sie_server.types.inputs import Item
@@ -201,7 +206,8 @@ def encode(
         Args:
             items: List of items to encode.
             output_types: Which outputs to compute (only "sparse" supported).
-            instruction: Optional instruction prefix.
+            instruction: Optional instruction prefix. For queries, ``None``
+                falls back to the ``default_instruction`` runtime option.
             is_query: Whether items are queries (affects template selection).
             prepared_items: Not used by this adapter.
 
@@ -221,7 +227,7 @@ def encode(
 
         texts = extract_texts(
             items,
-            instruction,
+            resolve_query_instruction(instruction, options, is_query=is_query),
             is_query=is_query,
             query_template=query_template,
             doc_template=doc_template,

packages/sie_server/src/sie_server/adapters/modernbert_flash/__init__.py

@@ -25,6 +25,7 @@
 from sie_server.adapters._utils import (
     extract_texts,
     resolve_embedding_options,
+    resolve_query_instruction,
     validate_output_types,
 )
 from sie_server.adapters.peft_lora_mixin import PEFTLoRAMixin
@@ -219,7 +220,8 @@ def encode(
         Args:
             items: List of items to encode.
             output_types: Which outputs to compute (only "dense" supported).
-            instruction: Optional instruction prefix.
+            instruction: Optional instruction prefix. For queries, ``None``
+                falls back to the ``default_instruction`` runtime option.
             is_query: Whether items are queries (affects template selection).
             prepared_items: Not used by this adapter.
             options: Runtime options for normalize, pooling, templates.
@@ -244,7 +246,7 @@ def encode(
 
         texts = extract_texts(
             items,
-            instruction,
+            resolve_query_instruction(instruction, options, is_query=is_query),
             is_query=is_query,
             query_template=query_template,
             doc_template=doc_template,

packages/sie_server/src/sie_server/adapters/nomic_flash/__init__.py

@@ -12,7 +12,12 @@
 from sie_server.adapters._flash_pack import build_position_ids, mean_pool_packed
 from sie_server.adapters._spec import AdapterSpec
 from sie_server.adapters._types import ERR_NOT_LOADED, ComputePrecision, PoolingStrategy
-from sie_server.adapters._utils import apply_rotary_pos_emb, extract_texts, validate_output_types
+from sie_server.adapters._utils import (
+    apply_rotary_pos_emb,
+    extract_texts,
+    resolve_query_instruction,
+    validate_output_types,
+)
 from sie_server.core.inference_output import EncodeOutput
 from sie_server.types.inputs import Item
 
@@ -236,7 +241,8 @@ def encode(
         Args:
             items: List of items to encode.
             output_types: Which outputs to compute (only "dense" supported).
-            instruction: Optional instruction prefix (unused, template-based).
+            instruction: Optional instruction prefix. For queries, ``None``
+                falls back to the ``default_instruction`` runtime option.
             is_query: Whether items are queries (affects template selection).
             prepared_items: Not used by this adapter.
 
@@ -256,7 +262,7 @@ def encode(
 
         texts = extract_texts(
             items,
-            instruction,
+            resolve_query_instruction(instruction, options, is_query=is_query),
             is_query=is_query,
             query_template=query_template,
             doc_template=doc_template,

packages/sie_server/src/sie_server/adapters/qwen2_flash/__init__.py

@@ -12,7 +12,12 @@
 from sie_server.adapters._flash_pack import build_position_ids, mean_pool_packed
 from sie_server.adapters._spec import AdapterSpec
 from sie_server.adapters._types import ERR_NOT_LOADED, ComputePrecision, PoolingStrategy
-from sie_server.adapters._utils import apply_rotary_pos_emb, extract_texts, validate_output_types
+from sie_server.adapters._utils import (
+    apply_rotary_pos_emb,
+    extract_texts,
+    resolve_query_instruction,
+    validate_output_types,
+)
 from sie_server.adapters.peft_lora_mixin import PEFTLoRAMixin
 from sie_server.core.inference_output import EncodeOutput
 from sie_server.types.inputs import Item
@@ -270,15 +275,10 @@ def encode(
         doc_template = opts.get("doc_template", self._doc_template)
         normalize = opts.get("normalize", self._normalize)
         pooling = opts.get("pooling", self._pooling)
-        # The profile's default_instruction is query-only and fills in only
-        # when the request gave no instruction; an explicit "" is kept.
-        effective_instruction = instruction
-        if effective_instruction is None and is_query:
-            effective_instruction = opts.get("default_instruction")
 
         texts = extract_texts(
             items,
-            effective_instruction,
+            resolve_query_instruction(instruction, opts, is_query=is_query),
             is_query=is_query,
             query_template=query_template,
             doc_template=doc_template,

packages/sie_server/src/sie_server/adapters/rope_flash/__init__.py

@@ -11,7 +11,12 @@
 from sie_server.adapters._flash_pack import mean_pool_packed
 from sie_server.adapters._spec import AdapterSpec
 from sie_server.adapters._types import ERR_NOT_LOADED, ComputePrecision, PoolingStrategy
-from sie_server.adapters._utils import apply_rotary_pos_emb, extract_texts, validate_output_types
+from sie_server.adapters._utils import (
+    apply_rotary_pos_emb,
+    extract_texts,
+    resolve_query_instruction,
+    validate_output_types,
+)
 from sie_server.adapters.peft_lora_mixin import PEFTLoRAMixin
 from sie_server.core.inference_output import EncodeOutput
 from sie_server.types.inputs import Item
@@ -172,7 +177,8 @@ def encode(
         Args:
             items: List of items to encode.
             output_types: Which outputs to compute (only "dense" supported).
-            instruction: Optional instruction prefix.
+            instruction: Optional instruction prefix. For queries, ``None``
+                falls back to the ``default_instruction`` runtime option.
             is_query: Whether items are queries (affects template selection).
             prepared_items: Not used by this adapter.
 
@@ -194,7 +200,7 @@ def encode(
 
         texts = extract_texts(
             items,
-            instruction,
+            resolve_query_instruction(instruction, options, is_query=is_query),
             is_query=is_query,
             query_template=query_template,
             doc_template=doc_template,

packages/sie_server/src/sie_server/adapters/splade_flash/adapter.py

@@ -14,7 +14,7 @@
 from sie_server.adapters._flash_pack import build_position_ids
 from sie_server.adapters._spec import AdapterSpec
 from sie_server.adapters._types import ERR_NOT_LOADED, ComputePrecision
-from sie_server.adapters._utils import extract_texts, validate_output_types
+from sie_server.adapters._utils import extract_texts, resolve_query_instruction, validate_output_types
 from sie_server.adapters.base import ModelAdapter
 from sie_server.adapters.peft_lora_mixin import PEFTLoRAMixin
 from sie_server.core.inference_output import EncodeOutput, SparseVector
@@ -230,7 +230,7 @@ def encode(
 
         texts = extract_texts(
             items,
-            instruction,
+            resolve_query_instruction(instruction, options, is_query=is_query),
             is_query=is_query,
             query_template=query_template,
             doc_template=doc_template,