feat(server): GLiClass task prompts, few-shot examples and label groups, plus instruct, Opir and multilingual models

README.md

@@ -47,7 +47,8 @@ One SIE cluster runs the inference behind a whole agent. Each task is a handful
 | **Document to markdown** | PDFs, Office files, and scans become clean markdown. | [`lightonocr`](packages/sie_server/models/lightonai__LightOnOCR-2-1B.yaml), [`glm-ocr`](packages/sie_server/models/zai-org__GLM-OCR.yaml), [`mineru`](packages/sie_server/models/opendatalab__MinerU2.5-Pro-2604-1.2B.yaml), [`paddleocr-vl`](packages/sie_server/models/PaddlePaddle__PaddleOCR-VL-1.5.yaml), [`docling`](packages/sie_server/models/docling.yaml) |
 | **Structured output** | Schema-valid JSON, extracted or generated. | [`gliner2`](packages/sie_server/models/fastino__gliner2-large-v1.yaml), [`nuner-zero`](packages/sie_server/models/numind__NuNER_Zero.yaml), [`qwen3.8-27b`](packages/sie_server/models/Qwen__Qwen3.8-27B-FP8.yaml), [`qwen3.6-27b`](packages/sie_server/models/Qwen__Qwen3.6-27B.yaml) |
 | **Decide** | Choice, yes/no, and score answers with probabilities to typed questions about a text or JSON state. | [`laya`](packages/sie_server/models/convaiinnovations__laya.yaml), [`laya-multilingual`](packages/sie_server/models/convaiinnovations__laya-multilingual.yaml), [`laya-typed-decisions`](packages/sie_server/models/convaiinnovations__laya-typed-decisions.yaml) |
-| **Guard content** | A Yes/No safety verdict, with the decision threshold tunable in the model config. | [`granite-guardian-2b`](packages/sie_server/models/ibm-granite__granite-guardian-3.0-2b.yaml) |
+| **Classify** | Zero-shot labels, with several label groups answered in one call. The instruct models also follow a task instruction and few-shot examples. | [`gliclass-large-v3`](packages/sie_server/models/knowledgator__gliclass-large-v3.0.yaml), [`gliclass-instruct-large`](packages/sie_server/models/knowledgator__gliclass-instruct-large-v1.0.yaml), [`gliclass-multilang-mini`](packages/sie_server/models/knowledgator__gliclass-multilang-mini.yaml) |
+| **Guard content** | A safety verdict: Yes/No with the threshold set in the model config, or safe/unsafe and policy-label scores with the threshold chosen per request. | [`granite-guardian-2b`](packages/sie_server/models/ibm-granite__granite-guardian-3.0-2b.yaml), [`opir-multitask-large`](packages/sie_server/models/knowledgator__opir-multitask-large-v1.0.yaml), [`opir-edge`](packages/sie_server/models/knowledgator__opir-edge-v1.0.yaml) |
 | **Run the agent loop** | Plan steps and call tools with an open LLM, streaming included. | [`qwen3.8-27b`](packages/sie_server/models/Qwen__Qwen3.8-27B-FP8.yaml), [`qwen3.6-27b`](packages/sie_server/models/Qwen__Qwen3.6-27B.yaml) |
 | **Translate** | Text between 400+ languages. | [`madlad400-3b-mt`](packages/sie_server/models/google__madlad400-3b-mt.yaml) |
 | **See images** | Caption, detect objects, and answer questions about images. | [`florence-2`](packages/sie_server/models/microsoft__Florence-2-large.yaml), [`owlv2`](packages/sie_server/models/google__owlv2-base-patch16-ensemble.yaml), [`grounding-dino`](packages/sie_server/models/IDEA-Research__grounding-dino-base.yaml) |

packages/sie_gateway/src/handlers/proxy.rs

@@ -374,6 +374,10 @@ const RESOURCE_EXHAUSTED_RETRY_AFTER: &str = RetryAfter::DEFAULT.resource_exhaus
 const LORA_LOADING_ERROR_CODE: &str = "LORA_LOADING";
 const LORA_LOADING_RETRY_AFTER: &str = RetryAfter::DEFAULT.lora_loading;
 const INVALID_INPUT_ERROR_CODE: &str = "INVALID_INPUT";
+/// Worker-side input exceeds the model's context window (for example a label
+/// set that does not fit). Caller-fixable, so it maps to 400 like
+/// ``INVALID_INPUT``; see ``sie_server.adapters.errors.InputTooLongError``.
+const INPUT_TOO_LONG_ERROR_CODE: &str = "INPUT_TOO_LONG";
 const PAYLOAD_TOO_LARGE_ERROR_CODE: &str = err_code::PAYLOAD_TOO_LARGE;
 
 /// Fallback `max_tokens` applied to a chat-completions request that
@@ -9002,6 +9006,7 @@ fn unanimous_terminal_client_error(
     let first = errors.first()?.error_code.as_deref()?;
     let (status, canonical) = match first {
         INVALID_INPUT_ERROR_CODE => (StatusCode::BAD_REQUEST, INVALID_INPUT_ERROR_CODE),
+        INPUT_TOO_LONG_ERROR_CODE => (StatusCode::BAD_REQUEST, INPUT_TOO_LONG_ERROR_CODE),
         PAYLOAD_TOO_LARGE_ERROR_CODE => {
             (StatusCode::PAYLOAD_TOO_LARGE, PAYLOAD_TOO_LARGE_ERROR_CODE)
         }
@@ -18468,6 +18473,7 @@ mod tests {
     fn test_unanimous_terminal_client_errors_map_to_400_and_413() {
         for (code, status) in [
             (INVALID_INPUT_ERROR_CODE, StatusCode::BAD_REQUEST),
+            (INPUT_TOO_LONG_ERROR_CODE, StatusCode::BAD_REQUEST),
             (PAYLOAD_TOO_LARGE_ERROR_CODE, StatusCode::PAYLOAD_TOO_LARGE),
         ] {
             let first = _err_result(Some(code), "rejected 1");
@@ -18485,6 +18491,27 @@ mod tests {
             unanimous_terminal_client_error(&[&invalid, &oversized]),
             None
         );
+        let too_long = _err_result(Some(INPUT_TOO_LONG_ERROR_CODE), "labels do not fit");
+        let failed = _err_result(Some("inference_error"), "backend failure");
+        assert_eq!(unanimous_terminal_client_error(&[&too_long, &failed]), None);
+    }
+
+    #[tokio::test]
+    async fn test_input_too_long_is_a_native_400_with_its_code() {
+        let too_long = _err_result(Some(INPUT_TOO_LONG_ERROR_CODE), "labels do not fit");
+        let (status, code) = unanimous_terminal_client_error(&[&too_long]).unwrap();
+        let response = build_terminal_client_error_response(status, code, "labels do not fit");
+        assert_eq!(response.status(), StatusCode::BAD_REQUEST);
+        assert_eq!(
+            response.headers().get("x-sie-error-code").unwrap(),
+            INPUT_TOO_LONG_ERROR_CODE
+        );
+        let body = axum::body::to_bytes(response.into_body(), 16 * 1024)
+            .await
+            .unwrap();
+        let value: serde_json::Value = serde_json::from_slice(&body).unwrap();
+        assert_eq!(value["detail"]["code"], INPUT_TOO_LONG_ERROR_CODE);
+        assert_eq!(value["detail"]["message"], "labels do not fit");
     }
 
     #[test]

packages/sie_sdk/README.md

@@ -34,6 +34,73 @@ for entry in scores["scores"]:
     print(entry["item_id"], entry["score"])
 ```
 
+## Zero-shot classification
+
+GLiClass models (`knowledgator/gliclass-*` and the `knowledgator/opir-*`
+guardrail models) score text against labels passed with each request. Every
+label comes back in `classifications`, sorted by score. By default the scores
+form one distribution that sums to 1; `options={"classification_type":
+"multi-label"}` scores each label independently instead.
+
+```python
+ticket = Item(text="I was charged twice and support has not replied in three days.")
+
+result = client.extract(
+    "knowledgator/gliclass-instruct-large-v1.0",
+    ticket,
+    labels=["billing", "bug report", "feature request"],
+    instruction="Classify the support ticket by its main topic.",
+)
+print(result["classifications"][0]["label"])  # billing
+```
+
+`instruction` is the model's task prompt; the `gliclass-instruct-*` and
+`opir-*` models are trained to follow one. Few-shot examples go in
+`options={"examples": [{"text": ..., "labels": [...]}]}`, with labels taken
+from the request's label set. They can move scores a long way, so check them
+on your own data. The model reads examples after the document, so a long
+document can push them out of the model's window; send
+`options={"overflow_policy": "truncate_text"}` to shorten the document instead.
+
+To answer several questions in one call, pass named `label_groups` instead of
+`labels`. Scores are normalized within each group, and `data` holds one answer
+per group:
+
+```python
+result = client.extract(
+    "knowledgator/gliclass-instruct-large-v1.0",
+    ticket,
+    instruction="Triage the support ticket.",
+    options={
+        "label_groups": {
+            "topic": ["billing", "bug report", "feature request"],
+            "urgency": ["low", "medium", "high"],
+            "needs_human": ["yes", "no"],
+        }
+    },
+)
+urgency = result["data"]["urgency"]
+print(urgency["choice"], urgency["confidence"])  # e.g. medium 0.78
+print(urgency["probabilities"])  # {"low": ..., "medium": ..., "high": ...}
+```
+
+Each group answers `{"type": "choice", "choice", "probabilities",
+"confidence"}`, where `confidence` is `1 - entropy / log(number of labels)`.
+With `"classification_type": "multi-label"` a group answers `{"labels",
+"probabilities"}`: every label scored independently, and `labels` lists those
+at or above `options.threshold` (0.5 when no threshold is set).
+`classifications` lists the same scores under `group.label` names. In a
+grouped request, an example's labels may be written as `"urgency.high"` or as
+`{"urgency": "high"}`.
+
+Usage counts each item's document tokens plus the tokens of the instruction
+and example texts sent with it, since the model encodes them for every item.
+Label names are not counted. With an instruction or examples, each item's
+count is capped at the model window minus the label prompt, unless the
+document count alone is already higher. An item whose document pushes the
+labels out of the window comes back with an `INPUT_TOO_LONG` error in its
+`error` field, and the other items still succeed.
+
 ## Generation prompts and guard verdicts
 
 `generate` and `stream_generate` treat text-only prompts as raw continuation

packages/sie_sdk/src/sie_sdk/types.py

@@ -511,7 +511,9 @@ class ExtractResult(TypedDict, total=False):
         relations: List of extracted relation triples.
         classifications: List of classification results.
         objects: List of detected objects with bounding boxes.
-        data: Additional structured extraction data (if output_schema was provided).
+        data: Structured extraction data: schema-driven results when
+            output_schema was provided, document parses, or one answer per
+            group when GLiClass options.label_groups was provided.
         error: Stable per-item failure when extraction did not complete.
         request: Request-scoped id, metered usage, and settled debit when
             supplied by the gateway.

packages/sie_server/README.md

@@ -48,6 +48,25 @@ removes matched stop sequences consistently from returned text, completion-token
 usage, and logprobs. Long completions delay the first visible text; existing
 request timeouts still apply. Other adapters are unaffected.
 
+### GLiClass usage
+
+For GLiClass classification, `usage.input_tokens` counts each item's document
+tokens plus the instruction and few-shot example texts sent with the request,
+because the model encodes that text again for every item. Label names,
+including the labels attached to examples, are not counted. With an
+instruction or examples, an item's count is capped at the model's
+`max_sequence_length` minus the label prompt, unless the document count alone
+is already higher. An item refused because its document pushes the labels out
+of the window returns a per-item `INPUT_TOO_LONG` error and counts nothing. A
+request that sends no instruction or examples is counted exactly as before.
+
+The instruction and each example text may be at most 2,048 characters, and
+together with the example labels at most 8,192 characters. Up to 32 examples
+are accepted, and they must leave room for the document in the model window.
+Label names are refused when their total length exceeds 16 characters per
+token of the window (8,192 characters for a 512-token model), more than any
+label prompt can fit.
+
 ## Configuration
 
 `sie-server` reads its config from `SIE_*` environment variables (Pydantic

packages/sie_server/bundles/default.yaml

@@ -79,8 +79,9 @@ deps:
   gliner2: '>=1.3.1,<2'
   # glirel
   glirel: '>=1.0,<2'
-  # gliclass
-  gliclass: '>=0.1,<1'
+  # gliclass (0.1.17: cross-attention scorer for the multilingual checkpoints;
+  # 0.1.18+ needs transformers 5)
+  gliclass: '>=0.1.17,<1'
   # gliner/glirel/gliclass shared dep
   loguru: '>=0.7,<1'
   # donut, florence2

packages/sie_server/models/knowledgator__gliclass-base-v3.0.yaml

@@ -0,0 +1,29 @@
+sie_id: knowledgator/gliclass-base-v3.0
+hf_id: knowledgator/gliclass-base-v3.0
+hf_revision: 77a70e6cd52e602ed18184ef37d18bdd3741e3d5
+inputs:
+  text: true
+  image: false
+  audio: false
+  video: false
+tasks:
+  encode: null
+  score: null
+  extract: {}
+max_sequence_length: 512
+profiles:
+  default:
+    max_batch_tokens: 16384
+    compute_precision: null
+    adapter_path: sie_server.adapters.gliclass:GLiClassAdapter
+    adapter_options:
+      loadtime:
+        classification_type: single-label
+      runtime:
+        threshold: 0.0
+  multi-label:
+    extends: default
+    adapter_options:
+      runtime:
+        classification_type: multi-label
+        threshold: 0.0

packages/sie_server/models/knowledgator__gliclass-edge-v3.0.yaml

@@ -0,0 +1,29 @@
+sie_id: knowledgator/gliclass-edge-v3.0
+hf_id: knowledgator/gliclass-edge-v3.0
+hf_revision: df03993a2ed98e5e4a0d2dd7efbbd105abe874cf
+inputs:
+  text: true
+  image: false
+  audio: false
+  video: false
+tasks:
+  encode: null
+  score: null
+  extract: {}
+max_sequence_length: 512
+profiles:
+  default:
+    max_batch_tokens: 16384
+    compute_precision: null
+    adapter_path: sie_server.adapters.gliclass:GLiClassAdapter
+    adapter_options:
+      loadtime:
+        classification_type: single-label
+      runtime:
+        threshold: 0.0
+  multi-label:
+    extends: default
+    adapter_options:
+      runtime:
+        classification_type: multi-label
+        threshold: 0.0

packages/sie_server/models/knowledgator__gliclass-instruct-base-v1.0.yaml

@@ -0,0 +1,29 @@
+sie_id: knowledgator/gliclass-instruct-base-v1.0
+hf_id: knowledgator/gliclass-instruct-base-v1.0
+hf_revision: 4f6a108b08a5537f395521d19b5073e197923dd3
+inputs:
+  text: true
+  image: false
+  audio: false
+  video: false
+tasks:
+  encode: null
+  score: null
+  extract: {}
+max_sequence_length: 512
+profiles:
+  default:
+    max_batch_tokens: 16384
+    compute_precision: null
+    adapter_path: sie_server.adapters.gliclass:GLiClassAdapter
+    adapter_options:
+      loadtime:
+        classification_type: single-label
+      runtime:
+        threshold: 0.0
+  multi-label:
+    extends: default
+    adapter_options:
+      runtime:
+        classification_type: multi-label
+        threshold: 0.0

packages/sie_server/models/knowledgator__gliclass-instruct-edge-v1.0.yaml

@@ -0,0 +1,29 @@
+sie_id: knowledgator/gliclass-instruct-edge-v1.0
+hf_id: knowledgator/gliclass-instruct-edge-v1.0
+hf_revision: 727be8a417f6a7718e591b025e07054c146d8139
+inputs:
+  text: true
+  image: false
+  audio: false
+  video: false
+tasks:
+  encode: null
+  score: null
+  extract: {}
+max_sequence_length: 512
+profiles:
+  default:
+    max_batch_tokens: 16384
+    compute_precision: null
+    adapter_path: sie_server.adapters.gliclass:GLiClassAdapter
+    adapter_options:
+      loadtime:
+        classification_type: single-label
+      runtime:
+        threshold: 0.0
+  multi-label:
+    extends: default
+    adapter_options:
+      runtime:
+        classification_type: multi-label
+        threshold: 0.0

packages/sie_server/models/knowledgator__gliclass-instruct-large-v1.0.yaml

@@ -0,0 +1,29 @@
+sie_id: knowledgator/gliclass-instruct-large-v1.0
+hf_id: knowledgator/gliclass-instruct-large-v1.0
+hf_revision: 825e5478c1bf4bffbf297690517097ccbdb2e006
+inputs:
+  text: true
+  image: false
+  audio: false
+  video: false
+tasks:
+  encode: null
+  score: null
+  extract: {}
+max_sequence_length: 512
+profiles:
+  default:
+    max_batch_tokens: 16384
+    compute_precision: null
+    adapter_path: sie_server.adapters.gliclass:GLiClassAdapter
+    adapter_options:
+      loadtime:
+        classification_type: single-label
+      runtime:
+        threshold: 0.0
+  multi-label:
+    extends: default
+    adapter_options:
+      runtime:
+        classification_type: multi-label
+        threshold: 0.0

packages/sie_server/models/knowledgator__gliclass-multilang-edge.yaml

@@ -0,0 +1,29 @@
+sie_id: knowledgator/gliclass-multilang-edge
+hf_id: knowledgator/gliclass-multilang-edge
+hf_revision: d16c08ef70547514081952104e6fc3d190d9ee39
+inputs:
+  text: true
+  image: false
+  audio: false
+  video: false
+tasks:
+  encode: null
+  score: null
+  extract: {}
+max_sequence_length: 512
+profiles:
+  default:
+    max_batch_tokens: 16384
+    compute_precision: null
+    adapter_path: sie_server.adapters.gliclass:GLiClassAdapter
+    adapter_options:
+      loadtime:
+        classification_type: single-label
+      runtime:
+        threshold: 0.0
+  multi-label:
+    extends: default
+    adapter_options:
+      runtime:
+        classification_type: multi-label
+        threshold: 0.0

packages/sie_server/models/knowledgator__gliclass-multilang-mini.yaml

@@ -0,0 +1,29 @@
+sie_id: knowledgator/gliclass-multilang-mini
+hf_id: knowledgator/gliclass-multilang-mini
+hf_revision: 0bd888b6c3ef9fca5f0a9d407bddfbbc7623486b
+inputs:
+  text: true
+  image: false
+  audio: false
+  video: false
+tasks:
+  encode: null
+  score: null
+  extract: {}
+max_sequence_length: 512
+profiles:
+  default:
+    max_batch_tokens: 16384
+    compute_precision: null
+    adapter_path: sie_server.adapters.gliclass:GLiClassAdapter
+    adapter_options:
+      loadtime:
+        classification_type: single-label
+      runtime:
+        threshold: 0.0
+  multi-label:
+    extends: default
+    adapter_options:
+      runtime:
+        classification_type: multi-label
+        threshold: 0.0