iatoolkit 2.42.2__tar.gz → 2.44.0__tar.gz

{iatoolkit-2.42.2/src/iatoolkit.egg-info → iatoolkit-2.44.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: iatoolkit
-Version: 2.42.2
+Version: 2.44.0
 Summary: IAToolkit
 Author: Fernando Libedinsky
 License-Expression: MIT
@@ -31,7 +31,7 @@ Requires-Dist: google-auth-httplib2==0.2.0
 Requires-Dist: google-auth-oauthlib==1.2.1
 Requires-Dist: google-cloud-core==2.4.1
 Requires-Dist: google-cloud-storage==3.0.0
-Requires-Dist: google-crc32c==1.6.0
+Requires-Dist: google-crc32c==1.8.0
 Requires-Dist: google-genai==1.57.0
 Requires-Dist: google-resumable-media==2.7.2
 Requires-Dist: googleapis-common-protos==1.66.0
@@ -88,6 +88,11 @@ Requires-Dist: Werkzeug==3.1.3
 Requires-Dist: pyjwt[crypto]>=2.8.0
 Provides-Extra: docling
 Requires-Dist: docling==2.72.0; extra == "docling"
+Provides-Extra: ocr
+Requires-Dist: pytesseract==0.3.13; extra == "ocr"
+Provides-Extra: inference
+Requires-Dist: torch==2.10.0; extra == "inference"
+Requires-Dist: diffusers==0.39.0; extra == "inference"
 Dynamic: license-file
 
 # 🧠 IAToolkit — Open-Source Framework for Real-World AI Assistants

iatoolkit-2.44.0/pyproject.toml

@@ -0,0 +1,64 @@
+[build-system]
+requires = ["setuptools>=68", "wheel", "build"]
+build-backend = "setuptools.build_meta"
+
+[project]
+name = "iatoolkit"
+dynamic = ["version", "dependencies"]
+requires-python = ">=3.12"
+description = "IAToolkit"
+readme = "readme.md"
+license = "MIT"
+authors = [{ name = "Fernando Libedinsky" }]
+
+# Capabilities that carry a heavy dependency and that most deployments never
+# turn on. Each of these was already being installed by hand into whatever
+# environment happened to need it, which is why a clean install could not run
+# the whole suite: nothing declared them, so nothing could install them on
+# purpose. Declared here, `pip install iatoolkit[ocr]` is a decision instead of
+# a memory.
+[project.optional-dependencies]
+# Parsing a document with the Docling pipeline rather than the basic provider.
+docling = [
+    "docling==2.72.0",
+]
+# Reading a scanned PDF. This extra is NOT enough to do OCR: pytesseract only
+# shells out to the tesseract binary, which pip cannot install. All three of
+# these have to be true, and `_get_tesseract_status()` checks them in order:
+#
+#   1. TESSERACT_ENABLED=true      — off by default, so this extra alone
+#                                    changes nothing at runtime
+#   2. this extra installed        — the only gate pip covers
+#   3. apt-get install tesseract-ocr tesseract-ocr-spa
+#                                  — the binary, and the Spanish data pack,
+#                                    because the call is image_to_string(lang="spa")
+#
+# Render's native Python runtime ships neither of the two system packages.
+# Installed in CI because the two OCR assertions in test_document_service.py
+# fake gates 1 and 3 but import pytesseract for real.
+ocr = [
+    "pytesseract==0.3.13",
+]
+# Running a model locally in `infra/inference_handler`, which imports both of
+# these at module level. Nothing in the source loads that module today — only
+# its own test does, by importlib — so treat these pins as untested: no code
+# path has exercised diffusers at any version.
+inference = [
+    "torch==2.10.0",
+    "diffusers==0.39.0",
+]
+
+[tool.setuptools]
+package-dir = {"" = "src"}
+
+[tool.setuptools.packages.find]
+where = ["src"]
+exclude = ["tests*", "*/tests*"]
+
+
+[tool.setuptools.dynamic]
+version = { attr = "iatoolkit.__version__" }
+dependencies = { file = ["requirements.txt"] }
+
+[tool.setuptools.package-data]
+iatoolkit = ["templates/**/*", "static/**/*", "locales/*.yaml", "config/*.yaml", "config/system_prompts/*.prompt"]

{iatoolkit-2.42.2 → iatoolkit-2.44.0}/requirements.txt

@@ -21,7 +21,9 @@ google-auth-httplib2==0.2.0
 google-auth-oauthlib==1.2.1
 google-cloud-core==2.4.1
 google-cloud-storage==3.0.0
-google-crc32c==1.6.0
+# 1.6.0 publishes no cp313 wheel. This was the only pin in the whole
+# dependency set that Python 3.13 required moving.
+google-crc32c==1.8.0
 google-genai==1.57.0
 google-resumable-media==2.7.2
 googleapis-common-protos==1.66.0

{iatoolkit-2.42.2 → iatoolkit-2.44.0}/src/iatoolkit/__init__.py

@@ -3,7 +3,7 @@
 #
 # IAToolkit is open source software.
 
-__version__ = "2.42.2"
+__version__ = "2.44.0"
 
 # Expose main classes and functions at the top level of the package
 

{iatoolkit-2.42.2 → iatoolkit-2.44.0}/src/iatoolkit/infra/inference_embeddings_client.py

@@ -37,5 +37,21 @@ class InferenceEmbeddingsClient:
         except Exception as e:
             raise RuntimeError(f"Error generando embedding con InferenceClient: {e}")
 
+    def get_embeddings(self, texts: List[str]) -> List[List[float]]:
+        """Embed a batch in one call. feature_extraction accepts a list and
+        returns one row per input; the single-text path already handled that shape."""
+        try:
+            embeddings = self.client.feature_extraction(list(texts))
+            if hasattr(embeddings, "tolist"):
+                embeddings = embeddings.tolist()
+            if not isinstance(embeddings, list) or len(embeddings) != len(texts):
+                raise RuntimeError(
+                    f"feature_extraction returned {len(embeddings) if isinstance(embeddings, list) else 'no'} "
+                    f"rows for {len(texts)} inputs."
+                )
+            return embeddings
+        except Exception as e:
+            raise RuntimeError(f"Error generando embeddings en lote con InferenceClient: {e}")
+
     def get_image_embedding(self, presigned_url: Optional[str] = None, image_bytes: Optional[bytes] = None) -> List[float]:
         raise NotImplementedError("Image embedding not implemented for this client.")

{iatoolkit-2.42.2 → iatoolkit-2.44.0}/src/iatoolkit/infra/inference_handler.py

@@ -192,12 +192,29 @@ class EndpointHandler:
         return {"embedding": vec}
 
     def _handle_text_embedding(self, inputs: dict) -> dict:
-        text = inputs.get("text")
-        if not isinstance(text, str) or not text.strip():
-            raise ValueError("Expected inputs.text for text embedding generation.")
+        """Embed one text or a batch, in the order received.
+
+        ``inputs.text`` (str) returns ``{"embedding": [...]}`` and ``inputs.texts``
+        (list) returns ``{"embeddings": [[...], ...]}``. Both shapes are answered so
+        an endpoint running this handler stays compatible with callers that predate
+        batching. The pooling below was already batch-shaped — it computed (B, D) and
+        returned row 0 — so the batch path is the same maths without the slice.
+        """
+        texts = inputs.get("texts")
+        batched = texts is not None
+        if batched:
+            if not isinstance(texts, list) or not texts:
+                raise ValueError("Expected inputs.texts to be a non-empty list of strings.")
+            if any(not isinstance(item, str) or not item.strip() for item in texts):
+                raise ValueError("Every entry in inputs.texts must be a non-empty string.")
+        else:
+            text = inputs.get("text")
+            if not isinstance(text, str) or not text.strip():
+                raise ValueError("Expected inputs.text for text embedding generation.")
+            texts = [text]
 
         encoded_input = self.processor_instance(
-            text,
+            texts,
             padding=True,
             truncation=True,
             return_tensors='pt'
@@ -213,7 +230,10 @@ class EndpointHandler:
             min=1e-9
         )
         sentence_embeddings = torch.nn.functional.normalize(sentence_embeddings, p=2, dim=1)
-        return {"embedding": sentence_embeddings[0].cpu().tolist()}
+        vectors = sentence_embeddings.cpu().tolist()
+        if batched:
+            return {"embeddings": vectors}
+        return {"embedding": vectors[0]}
 
     # Backward-compatible alias: existing tests/callers may still refer to _handle_minilm
     def _handle_minilm(self, inputs: dict) -> dict:

{iatoolkit-2.42.2 → iatoolkit-2.44.0}/src/iatoolkit/infra/jina_embeddings_client.py

@@ -36,6 +36,23 @@ class JinaEmbeddingsClient:
         response = self._post(payload)
         return response["data"][0]["embedding"]
 
+    def get_embeddings(self, texts: List[str]) -> List[List[float]]:
+        """Embed a batch in one request. The API already took a list; this uses it."""
+        payload = {
+            "model": self.model,
+            "normalized": self.normalized,
+            "embedding_type": "float",
+            "input": list(texts),
+        }
+        response = self._post(payload)
+        data = response["data"]
+        if len(data) != len(texts):
+            raise ValueError(f"Jina returned {len(data)} embeddings for {len(texts)} inputs.")
+        # Sort by the API's own index rather than trusting arrival order: a silent
+        # reordering would attach each vector to the wrong text.
+        ordered = sorted(data, key=lambda item: item.get("index", 0))
+        return [item["embedding"] for item in ordered]
+
     def get_image_embedding(self,
                             presigned_url: Optional[str] = None,
                             image_bytes: Optional[bytes] = None

{iatoolkit-2.42.2 → iatoolkit-2.44.0}/src/iatoolkit/services/embedding_service.py

@@ -34,6 +34,22 @@ class EmbeddingClientWrapper:
         """Generates and returns an embedding for the given text."""
         raise NotImplementedError
 
+    def get_embeddings(self, texts: list[str], suppress_error_logging: bool = False) -> list[list[float]]:
+        """Embeds several texts, returning one vector per input in the same order.
+
+        The default walks the single-text path so every provider works without
+        changes; wrappers whose backend accepts a list override this with one
+        request per batch. Callers must not assume a single round trip.
+        """
+        return [
+            self.get_embedding(text, suppress_error_logging=suppress_error_logging)
+            for text in texts
+        ]
+
+    def supports_batching(self) -> bool:
+        """Whether get_embeddings issues one request per batch instead of per text."""
+        return False
+
     def get_image_embedding(self,
                             presigned_url: Optional[str] = None,
                             image_bytes: Optional[bytes] = None
@@ -71,6 +87,39 @@ class HuggingFaceClientWrapper(EmbeddingClientWrapper):
         )
         return result["embedding"]
 
+    def supports_batching(self) -> bool:
+        # Opt-in per tool: an endpoint still running a handler that predates
+        # inputs.texts would reject the batch payload, so this stays off until the
+        # deployment is confirmed via `supports_batch_embedding: true`.
+        return bool(self._tool_config().get("supports_batch_embedding"))
+
+    def _tool_config(self) -> dict:
+        try:
+            return self.inference_service._get_tool_config(self.company_short_name, self.tool_name) or {}
+        except Exception:
+            logging.debug("Could not read inference tool config for batching.", exc_info=True)
+            return {}
+
+    def get_embeddings(self, texts: list[str], suppress_error_logging: bool = False) -> list[list[float]]:
+        if not self.supports_batching():
+            return super().get_embeddings(texts, suppress_error_logging=suppress_error_logging)
+
+        result = self.inference_service.predict(
+            self.company_short_name,
+            self.tool_name,
+            {"mode": "text", "texts": list(texts)},
+            suppress_error_logging=suppress_error_logging,
+        )
+        embeddings = result.get("embeddings")
+        if not isinstance(embeddings, list) or len(embeddings) != len(texts):
+            # A silent length mismatch would pair each text with the wrong vector,
+            # which no downstream check would catch. Fail instead.
+            raise ValueError(
+                f"Batch embedding returned {len(embeddings) if isinstance(embeddings, list) else 'no'} "
+                f"vectors for {len(texts)} inputs."
+            )
+        return embeddings
+
     def get_image_embedding(self,
                             presigned_url: Optional[str] = None,
                             image_bytes: Optional[bytes] = None
@@ -129,6 +178,24 @@ class OpenAIClientWrapper(EmbeddingClientWrapper):
         response = self.client.embeddings.create(**kwargs)
         return response.data[0].embedding
 
+    def supports_batching(self) -> bool:
+        return True
+
+    def get_embeddings(self, texts: list[str], suppress_error_logging: bool = False) -> list[list[float]]:
+        kwargs = {
+            "input": [text.replace("\n", " ") for text in texts],
+            "model": self.model,
+        }
+        if self.dimensions is not None:
+            kwargs["dimensions"] = self.dimensions
+        response = self.client.embeddings.create(**kwargs)
+        # The API documents input order, but pairing the wrong vector to a text is
+        # invisible downstream, so sort by index rather than trusting arrival order.
+        ordered = sorted(response.data, key=lambda item: item.index)
+        if len(ordered) != len(texts):
+            raise ValueError(f"Batch embedding returned {len(ordered)} vectors for {len(texts)} inputs.")
+        return [item.embedding for item in ordered]
+
 class CustomClassClientWrapper(EmbeddingClientWrapper):
     """
     Adapter for custom embedding classes defined by the user.
@@ -151,6 +218,20 @@ class CustomClassClientWrapper(EmbeddingClientWrapper):
             return embedding[0]
         return embedding
 
+    def supports_batching(self) -> bool:
+        return hasattr(self.client, "get_embeddings")
+
+    def get_embeddings(self, texts: list[str], suppress_error_logging: bool = False) -> list[list[float]]:
+        if not self.supports_batching():
+            return super().get_embeddings(texts, suppress_error_logging=suppress_error_logging)
+        embeddings = self.client.get_embeddings(list(texts))
+        if not isinstance(embeddings, list) or len(embeddings) != len(texts):
+            raise ValueError(
+                f"Batch embedding returned {len(embeddings) if isinstance(embeddings, list) else 'no'} "
+                f"vectors for {len(texts)} inputs."
+            )
+        return embeddings
+
     def get_image_embedding(self,
                             presigned_url: Optional[str] = None,
                             image_bytes: Optional[bytes] = None
@@ -504,6 +585,80 @@ class EmbeddingService:
                 logging.error(f"Error generating embedding for text: {text[:80]}... - {e}")
             raise
 
+    # Batch defaults: 32 texts is a good round trip for short records, and the
+    # character budget guards against a few long findings blowing the payload.
+    DEFAULT_EMBED_BATCH_SIZE = 32
+    DEFAULT_EMBED_BATCH_MAX_CHARS = 40000
+
+    def embed_texts(
+            self,
+            company_short_name: str,
+            texts: list[str],
+            model_type: str = 'text',
+            batch_size: Optional[int] = None,
+            max_chars_per_batch: Optional[int] = None,
+            suppress_error_logging: bool = False,
+    ) -> list[list[float]]:
+        """Embed many texts, returning one vector per input in the same order.
+
+        Resolves the company and the client once instead of once per text, and
+        sends batches when the provider supports them. If a batch fails, its
+        texts are retried one by one so a single bad input cannot lose the rest.
+        """
+        if not texts:
+            return []
+
+        company = self.profile_repo.get_company_by_short_name(company_short_name)
+        if not company:
+            raise ValueError(self.i18n_service.t('errors.company_not_found', company_short_name=company_short_name))
+        client_wrapper = self.client_factory.get_client(company_short_name, model_type)
+
+        size = int(batch_size or self.DEFAULT_EMBED_BATCH_SIZE)
+        char_budget = int(max_chars_per_batch or self.DEFAULT_EMBED_BATCH_MAX_CHARS)
+        embeddings: list[list[float]] = []
+        for batch in self._chunk_texts(texts, size=size, char_budget=char_budget):
+            embeddings.extend(
+                self._embed_batch(
+                    client_wrapper,
+                    batch,
+                    suppress_error_logging=suppress_error_logging,
+                )
+            )
+        if len(embeddings) != len(texts):
+            raise ValueError(f"Embedded {len(embeddings)} vectors for {len(texts)} texts.")
+        return embeddings
+
+    def _embed_batch(self, client_wrapper, batch: list[str], *, suppress_error_logging: bool) -> list[list[float]]:
+        try:
+            return client_wrapper.get_embeddings(batch, suppress_error_logging=suppress_error_logging)
+        except Exception as error:
+            if len(batch) == 1 or not client_wrapper.supports_batching():
+                raise
+            logging.warning(
+                "Batch embedding of %s texts failed (%s); retrying them individually.",
+                len(batch),
+                error,
+            )
+            return [
+                client_wrapper.get_embedding(text, suppress_error_logging=suppress_error_logging)
+                for text in batch
+            ]
+
+    @staticmethod
+    def _chunk_texts(texts: list[str], *, size: int, char_budget: int):
+        """Split by count and by characters, so a few long texts do not oversize a batch."""
+        batch: list[str] = []
+        batch_chars = 0
+        for text in texts:
+            length = len(text or "")
+            if batch and (len(batch) >= size or batch_chars + length > char_budget):
+                yield batch
+                batch, batch_chars = [], 0
+            batch.append(text)
+            batch_chars += length
+        if batch:
+            yield batch
+
     def embed_image(self, company_short_name: str,
                     presigned_url: Optional[str] = None,
                     image_bytes: Optional[bytes] = None) -> list[float]:

{iatoolkit-2.42.2 → iatoolkit-2.44.0/src/iatoolkit.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: iatoolkit
-Version: 2.42.2
+Version: 2.44.0
 Summary: IAToolkit
 Author: Fernando Libedinsky
 License-Expression: MIT
@@ -31,7 +31,7 @@ Requires-Dist: google-auth-httplib2==0.2.0
 Requires-Dist: google-auth-oauthlib==1.2.1
 Requires-Dist: google-cloud-core==2.4.1
 Requires-Dist: google-cloud-storage==3.0.0
-Requires-Dist: google-crc32c==1.6.0
+Requires-Dist: google-crc32c==1.8.0
 Requires-Dist: google-genai==1.57.0
 Requires-Dist: google-resumable-media==2.7.2
 Requires-Dist: googleapis-common-protos==1.66.0
@@ -88,6 +88,11 @@ Requires-Dist: Werkzeug==3.1.3
 Requires-Dist: pyjwt[crypto]>=2.8.0
 Provides-Extra: docling
 Requires-Dist: docling==2.72.0; extra == "docling"
+Provides-Extra: ocr
+Requires-Dist: pytesseract==0.3.13; extra == "ocr"
+Provides-Extra: inference
+Requires-Dist: torch==2.10.0; extra == "inference"
+Requires-Dist: diffusers==0.39.0; extra == "inference"
 Dynamic: license-file
 
 # 🧠 IAToolkit — Open-Source Framework for Real-World AI Assistants

{iatoolkit-2.42.2 → iatoolkit-2.44.0}/src/iatoolkit.egg-info/requires.txt

@@ -21,7 +21,7 @@ google-auth-httplib2==0.2.0
 google-auth-oauthlib==1.2.1
 google-cloud-core==2.4.1
 google-cloud-storage==3.0.0
-google-crc32c==1.6.0
+google-crc32c==1.8.0
 google-genai==1.57.0
 google-resumable-media==2.7.2
 googleapis-common-protos==1.66.0
@@ -79,3 +79,10 @@ pyjwt[crypto]>=2.8.0
 
 [docling]
 docling==2.72.0
+
+[inference]
+torch==2.10.0
+diffusers==0.39.0
+
+[ocr]
+pytesseract==0.3.13

iatoolkit-2.42.2/pyproject.toml

@@ -1,32 +0,0 @@
-[build-system]
-requires = ["setuptools>=68", "wheel", "build"]
-build-backend = "setuptools.build_meta"
-
-[project]
-name = "iatoolkit"
-dynamic = ["version", "dependencies"]
-requires-python = ">=3.12"
-description = "IAToolkit"
-readme = "readme.md"
-license = "MIT"
-authors = [{ name = "Fernando Libedinsky" }]
-
-[project.optional-dependencies]
-docling = [
-    "docling==2.72.0",
-]
-
-[tool.setuptools]
-package-dir = {"" = "src"}
-
-[tool.setuptools.packages.find]
-where = ["src"]
-exclude = ["tests*", "*/tests*"]
-
-
-[tool.setuptools.dynamic]
-version = { attr = "iatoolkit.__version__" }
-dependencies = { file = ["requirements.txt"] }
-
-[tool.setuptools.package-data]
-iatoolkit = ["templates/**/*", "static/**/*", "locales/*.yaml", "config/*.yaml", "config/system_prompts/*.prompt"]