banks 2.3.0__py3-none-any.whl → 2.4.0__py3-none-any.whl

banks/__about__.py

@@ -1,4 +1,4 @@
 # SPDX-FileCopyrightText: 2023-present Massimiliano Pippi <mpippi@gmail.com>
 #
 # SPDX-License-Identifier: MIT
-__version__ = "2.3.0"
+__version__ = "2.4.0"

banks/filters/audio.py

@@ -6,7 +6,9 @@ from pathlib import Path
 from typing import cast
 from urllib.parse import urlparse
 
-from banks.types import AudioFormat, ContentBlock, InputAudio
+import filetype  # type: ignore[import-untyped]
+
+from banks.types import AudioFormat, ContentBlock, InputAudio, resolve_binary
 
 BASE64_AUDIO_REGEX = re.compile(r"audio\/.*;base64,.*")
 
@@ -38,7 +40,18 @@ def _get_audio_format_from_url(url: str) -> AudioFormat:
     return "mp3"
 
 
-def audio(value: str) -> str:
+def _get_audio_format_from_bytes(data: bytes) -> AudioFormat:
+    """Extract audio format from bytes data using filetype library."""
+    kind = filetype.guess(data)
+    if kind is not None:
+        fmt = kind.extension
+        if fmt in ("mp3", "wav", "m4a", "webm", "ogg", "flac"):
+            return cast(AudioFormat, fmt)
+    # Default to mp3 if format cannot be determined
+    return "mp3"
+
+
+def audio(value: str | bytes) -> str:
     """Wrap the filtered value into a ContentBlock of type audio.
 
     The resulting ChatMessage will have the field `content` populated with a list of ContentBlock objects.
@@ -51,7 +64,10 @@ def audio(value: str) -> str:
         {{ "https://example.com/audio.mp3" | audio }}
         ```
     """
-    if _is_url(value):
+    if isinstance(value, bytes):
+        audio_format = _get_audio_format_from_bytes(resolve_binary(value, as_base64=False))
+        input_audio = InputAudio.from_bytes(value, audio_format=audio_format)
+    elif _is_url(value):
         audio_format = _get_audio_format_from_url(value)
         input_audio = InputAudio.from_url(value, audio_format)
     else:

banks/filters/document.py

@@ -1,12 +1,15 @@
 # SPDX-FileCopyrightText: 2023-present Massimiliano Pippi <mpippi@gmail.com>
 #
 # SPDX-License-Identifier: MIT
+import mimetypes
 import re
 from pathlib import Path
 from typing import cast
 from urllib.parse import urlparse
 
-from banks.types import ContentBlock, DocumentFormat, InputDocument
+import filetype  # type: ignore[import-untyped]
+
+from banks.types import ContentBlock, DocumentFormat, InputDocument, resolve_binary
 
 BASE64_DOCUMENT_REGEX = re.compile(r"(text|application)\/.*;base64,.*")
 
@@ -36,7 +39,7 @@ def _get_document_format_from_url(url: str) -> DocumentFormat:
     # text/css
     # text/plain
     # text/xml
-    # text/scv
+    # text/csv
     # text/rtf
     # text/javascript
     # application/json
@@ -68,13 +71,46 @@ def _get_document_format_from_url(url: str) -> DocumentFormat:
         "javascript",
         "json",
     ):
+        # Because Claude only supports pdf and text, and Gemini only supports a small subset of text formats,
+        # we can default to 'txt' for any text-based format that is not pdf. This allows the data to be sent to the llm
+        # in an acceptable format, but the LLM should still be able to understand the content: e.g., html, markdown,
+        # xml, etc.
         if path.endswith(f".{fmt}"):
+            if fmt == "pdf":
+                return cast(DocumentFormat, "pdf")
+            return "txt"
+    mime = mimetypes.guess_type(path)[0]
+    if mime is not None and mime.startswith("text/"):
+        return "txt"
+    # With urls, the likelihood seems sufficiently high that it's probably a pdf if not otherwise indicated
+    if mime is None:
+        return "pdf"
+    # Document type indicated to be other than pdf or text type
+    raise ValueError("Unsupported document format: " + path)
+
+
+def _get_document_format_from_bytes(data: bytes) -> DocumentFormat:
+    """Extract document format from bytes data using filetype library."""
+    # First check for pdf (only non text based format) and RTF formats (can be detected by file header)
+    kind = filetype.guess(data)
+    if kind is not None:
+        fmt = kind.extension
+        if fmt == "pdf":
             return cast(DocumentFormat, fmt)
-    # Default to pdf if format cannot be determined
-    return "pdf"
+
+    # filetype is good at detecting binary formats, but not text-based ones.
+    # So, this is a good indicator that it's text-based.
+    # Because Claude only supports pdf and text, and Gemini only supports a small subset of text formats,
+    # we can default to 'txt' for any text-based format that is not pdf. This allows the data to be sent to the llm in
+    # an acceptable format, but the LLM should still be able to understand the content: e.g., html, markdown, xml, etc.
+    # If detecting text types should become desirable, I recommend using something like Google magicka
+    if kind is None or kind.extension == "rtf":
+        return "txt"
+    # There are many common document types (like word, excel, powerpoint, etc.) that are not supported.
+    raise ValueError("Unsupported document format: " + kind.extension)
 
 
-def document(value: str) -> str:
+def document(value: str | bytes) -> str:
     """Wrap the filtered value into a ContentBlock of type document.
 
     The resulting ChatMessage will have the field `content` populated with a list of ContentBlock objects.
@@ -87,7 +123,10 @@ def document(value: str) -> str:
         {{ "https://example.com/document.pdf" | document }}
         ```
     """
-    if _is_url(value):
+    if isinstance(value, bytes):
+        document_format = _get_document_format_from_bytes(resolve_binary(value, as_base64=False))
+        input_document = InputDocument.from_bytes(value, document_format=document_format)
+    elif _is_url(value):
         document_format = _get_document_format_from_url(value)
         input_document = InputDocument.from_url(value, document_format)
     else:

banks/filters/image.py

@@ -22,7 +22,7 @@ def _is_url(string: str) -> bool:
     return True
 
 
-def image(value: str) -> str:
+def image(value: str | bytes) -> str:
     """Wrap the filtered value into a ContentBlock of type image.
 
     The resulting ChatMessage will have the field `content` populated with a list of ContentBlock objects.
@@ -38,7 +38,9 @@ def image(value: str) -> str:
         this filter marks the content to cache by surrounding it with `<content_block>` and
         `</content_block>`, so it's only useful when used within a `{% chat %}` block.
     """
-    if _is_url(value):
+    if isinstance(value, bytes):
+        image_url = ImageUrl.from_bytes(bytes_str=value)
+    elif _is_url(value):
         image_url = ImageUrl(url=value)
     else:
         image_url = ImageUrl.from_path(Path(value))

banks/filters/video.py

@@ -6,11 +6,39 @@ from pathlib import Path
 from typing import cast
 from urllib.parse import urlparse
 
-from banks.types import ContentBlock, InputVideo, VideoFormat
+import filetype  # type: ignore[import-untyped]
+from filetype.types.video import IsoBmff  # type: ignore[import-untyped]
+
+from banks.types import ContentBlock, InputVideo, VideoFormat, resolve_binary
 
 BASE64_VIDEO_REGEX = re.compile(r"video\/.*;base64,.*")
 
 
+class M3gp(IsoBmff):
+    """
+    Implements the 3gp video type matcher.
+
+    The type matcher in the filetype lib does not work correctly for 3gp files,
+    so implement our own here.
+    """
+
+    MIME = "video/3gpp"
+    EXTENSION = "3gp"
+
+    def __init__(self):
+        super().__init__(mime=M3gp.MIME, extension=M3gp.EXTENSION)
+
+    def match(self, buf):
+        if not self._is_isobmff(buf):
+            return False
+
+        major_brand, _, compatible_brands = self._get_ftyp(buf)
+        for brand in compatible_brands:
+            if brand in ["3gp4", "3gp5", "3gpp"]:
+                return True
+        return major_brand in ["3gp4", "3gp5", "3gpp"]
+
+
 def _is_url(string: str) -> bool:
     """Check if a string is a URL."""
     result = urlparse(string)
@@ -40,7 +68,22 @@ def _get_video_format_from_url(url: str) -> VideoFormat:
     return "mp4"
 
 
-def video(value: str) -> str:
+def _get_video_format_from_bytes(data: bytes) -> VideoFormat:
+    """Extract video format from bytes data using filetype library."""
+    m3gp = M3gp()
+    if m3gp not in filetype.types:
+        filetype.add_type(m3gp)
+
+    kind = filetype.guess(data)
+    if kind is not None:
+        fmt = kind.extension
+        if fmt in ("mp4", "mpg", "mov", "avi", "flv", "webm", "wmv", "3gp"):
+            return cast(VideoFormat, fmt)
+    # Default to mp4 if format cannot be determined
+    return "mp4"
+
+
+def video(value: str | bytes) -> str:
     """Wrap the filtered value into a ContentBlock of type video.
 
     The resulting ChatMessage will have the field `content` populated with a list of ContentBlock objects.
@@ -53,7 +96,10 @@ def video(value: str) -> str:
         {{ "https://example.com/video.mp4" | video }}
         ```
     """
-    if _is_url(value):
+    if isinstance(value, bytes):
+        video_format = _get_video_format_from_bytes(resolve_binary(value, as_base64=False))
+        input_video = InputVideo.from_bytes(value, video_format=video_format)
+    elif _is_url(value):
         video_format = _get_video_format_from_url(value)
         input_video = InputVideo.from_url(value, video_format)
     else:

banks/types.py

@@ -5,11 +5,14 @@ from __future__ import annotations
 
 import base64
 import re
+from base64 import b64decode, b64encode
+from binascii import Error as BinasciiError
 from enum import Enum
 from inspect import Parameter, getdoc, signature
 from pathlib import Path
 from typing import Callable, Literal, Union, cast
 
+import filetype  # type: ignore[import-untyped]
 from pydantic import BaseModel
 from typing_extensions import Self
 
@@ -19,6 +22,31 @@ from .utils import parse_params_from_docstring, python_type_to_jsonschema
 CONTENT_BLOCK_REGEX = re.compile(r"(<content_block>\{.*?\}<\/content_block>)|([^<](?:(?!<content_block>)[\s\S])*)")
 
 
+def resolve_binary(bytes_str: bytes, *, as_base64: bool = True) -> bytes:
+    """
+    Resolve binary data between base64 and raw bytes.
+
+    Args:
+        bytes_str: Bytes data
+        as_base64: Whether to return base64 encoded bytes or raw bytes
+
+    Returns:
+        b64 encoded bytes if input is not base64 encoded, else returns input as is.
+    """
+    # check if bytes_str is base64 encoded
+    try:
+        # Check if raw_bytes is already base64 encoded.
+        # b64decode() can succeed on random binary data, so we
+        # pass verify=True to make sure it's not a false positive
+        raw_bytes = base64.b64decode(bytes_str, validate=True)
+        b64_bytes = bytes_str
+    except BinasciiError:
+        # b64decode failed, leave as is
+        raw_bytes = bytes_str
+        b64_bytes = b64encode(bytes_str)
+    return b64_bytes if as_base64 else raw_bytes
+
+
 class ContentBlockType(str, Enum):
     text = "text"
     image_url = "image_url"
@@ -34,6 +62,14 @@ class CacheControl(BaseModel):
 class ImageUrl(BaseModel):
     url: str
 
+    @staticmethod
+    def _mimetype_from_bytes(raw_bytes: bytes) -> str:
+        kind = filetype.guess(raw_bytes)
+        if kind is not None:
+            return kind.mime
+        # Default to jpeg if format cannot be determined
+        return "image/jpeg"
+
     @classmethod
     def from_base64(cls, media_type: str, base64_str: str) -> Self:
         return cls(url=f"data:{media_type};base64,{base64_str}")
@@ -41,12 +77,31 @@ class ImageUrl(BaseModel):
     @classmethod
     def from_path(cls, file_path: Path) -> Self:
         with open(file_path, "rb") as image_file:
-            return cls.from_base64("image/jpeg", base64.b64encode(image_file.read()).decode("utf-8"))
+            raw_bytes = image_file.read()
+            mimetype = cls._mimetype_from_bytes(raw_bytes)
+            return cls.from_base64(mimetype, base64.b64encode(raw_bytes).decode("utf-8"))
+
+    @classmethod
+    def from_bytes(cls, bytes_str: bytes) -> Self:
+        """Create ImageUrl from bytes
+        Args:
+            bytes_str: Bytes data
+        Returns:
+            ImageUrl instance with base64 encoded bytes as URL
+        """
+        b64_bytes = resolve_binary(bytes_str)
+        mimetype = cls._mimetype_from_bytes(b64decode(b64_bytes))
+        return cls.from_base64(mimetype, b64_bytes.decode("utf-8"))
 
 
 AudioFormat = Literal["mp3", "wav", "m4a", "webm", "ogg", "flac"]
-VideoFormat = Literal["mp4", "mpeg", "mov", "avi", "flv", "mpg", "webm", "wmv", "3gpp"]
-DocumentFormat = Literal["pdf", "html", "css", "plain", "xml", "csv", "rtf", "javascript", "json"]
+VideoFormat = Literal["mp4", "mpg", "mov", "avi", "flv", "webm", "wmv", "3gp", "3gpp"]
+# Because Claude only supports pdf and text, and Gemini only supports a small subset of text formats,
+# we can default to 'txt' for any text-based format that is not pdf. This allows the data to be sent to the llm
+# in an acceptable format, but the LLM should still be able to understand the content: e.g., html, markdown,
+# xml, etc.
+# If detecting text types should become desirable, I recommend using something like Google magicka
+DocumentFormat = Literal["pdf", "txt"]
 
 
 class InputAudio(BaseModel):
@@ -73,6 +128,21 @@ class InputAudio(BaseModel):
         """
         return cls(data=url, format=audio_format)
 
+    @classmethod
+    def from_bytes(cls, bytes_str: bytes, audio_format: AudioFormat) -> Self:
+        """Create InputAudio from bytes
+
+        Args:
+            bytes_str: Bytes data
+            audio_format: The audio format
+
+        Returns:
+            InputAudio instance with base64 encoded bytes as data
+        """
+        b64_bytes = resolve_binary(bytes_str)
+        encoded_str = b64_bytes.decode("utf-8")
+        return cls(data=encoded_str, format=audio_format)
+
 
 class InputVideo(BaseModel):
     data: str
@@ -98,6 +168,21 @@ class InputVideo(BaseModel):
         """
         return cls(data=url, format=video_format)
 
+    @classmethod
+    def from_bytes(cls, bytes_str: bytes, video_format: VideoFormat) -> Self:
+        """Create InputVideo from bytes
+
+        Args:
+            bytes_str: Bytes data
+            video_format: The video format
+
+        Returns:
+            InputVideo instance with base64 encoded bytes as data
+        """
+        b64_bytes = resolve_binary(bytes_str)
+        encoded_str = b64_bytes.decode("utf-8")
+        return cls(data=encoded_str, format=video_format)
+
 
 class InputDocument(BaseModel):
     data: str
@@ -123,6 +208,21 @@ class InputDocument(BaseModel):
         """
         return cls(data=url, format=document_format)
 
+    @classmethod
+    def from_bytes(cls, bytes_str: bytes, document_format: DocumentFormat) -> Self:
+        """Create InputDocument from bytes
+
+        Args:
+            bytes_str: Bytes data
+            document_format: The document format
+
+        Returns:
+            InputDocument instance with base64 encoded bytes as data
+        """
+        b64_bytes = resolve_binary(bytes_str)
+        encoded_str = b64_bytes.decode("utf-8")
+        return cls(data=encoded_str, format=document_format)
+
 
 class ContentBlock(BaseModel):
     type: ContentBlockType

{banks-2.3.0.dist-info → banks-2.4.0.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: banks
-Version: 2.3.0
+Version: 2.4.0
 Summary: A prompt programming language
 Project-URL: Documentation, https://github.com/masci/banks#readme
 Project-URL: Issues, https://github.com/masci/banks/issues
@@ -21,6 +21,7 @@ Classifier: Programming Language :: Python :: Implementation :: PyPy
 Requires-Python: >=3.9
 Requires-Dist: deprecated
 Requires-Dist: eval-type-backport; python_version < '3.10'
+Requires-Dist: filetype>=1.2.0
 Requires-Dist: griffe
 Requires-Dist: jinja2
 Requires-Dist: platformdirs

{banks-2.3.0.dist-info → banks-2.4.0.dist-info}/RECORD

@@ -1,30 +1,30 @@
-banks/__about__.py,sha256=k_flpZerbrh6w78dg4-Rmizti2D8Fd5uyalbWC-tEZg,132
+banks/__about__.py,sha256=Kbqara7VPF_4Txd0QVMWDNL5EqtjtYKEQFgzvx2Dqgc,132
 banks/__init__.py,sha256=4IBopxXstFZliCvSjOuTurSQb32Vy26EXOPhmNZ4Hus,334
 banks/cache.py,sha256=uUGAu82-mfrscc2q24x19ZMZBkoQzf3hh7_V300J-Ik,1069
 banks/config.py,sha256=Ry2pdba1pgepsfmj41hxMdfQCT9pMjdV6TRWySpMiSY,1192
 banks/env.py,sha256=oGC4wjmF0-NTwoi49NooCLqqT44htn6EWDnHsDl2n0I,1347
 banks/errors.py,sha256=EnKRBhHmo8KEcSg3YoDBtVEaooJj9uSqRV1wnzUtrJU,580
 banks/prompt.py,sha256=LXaDGHBywFkD_JmWPQOXqyuQKZgozzDIb5lSZ12xf5A,8505
-banks/types.py,sha256=34YzR2Dzf7gpDlpODv-pHFZsGbNxt5UYWNxkih1wo1U,7747
+banks/types.py,sha256=p74ZUJfH0vZnneOEsAmiHoDGfsbjNLO-VDHYPuov2Yo,11347
 banks/utils.py,sha256=ZetGG3qhXMYOitDZQCWbE33wHEqR0ih2ZEg_dIW8OeI,1827
 banks/extensions/__init__.py,sha256=Lx4UrOzywYQY7a8qvIqvc3ql54nwK0lNP7x3jYdbREY,110
 banks/extensions/chat.py,sha256=VV6UV1wQZcJ0KbIFHSFmDeptWtww4o2IXF5pXB6TpTM,2478
 banks/extensions/completion.py,sha256=p6NdzA5kOuWZ0BIcGQH86Ji4Z4PFz0-h_G2cHgKdYvw,7861
 banks/extensions/docs.py,sha256=vWOZvu2JoS4LwUG-BR3jPqThirYvu3Fdba331UxooYM,1098
 banks/filters/__init__.py,sha256=fcAlKqgDSX19JDQHfTTOtpotxbdC84QbcFF3dKTtEog,430
-banks/filters/audio.py,sha256=H_kOKhVKk1BdKmVgia682jn_09FInGjehEtXuUtXZ6Y,1900
+banks/filters/audio.py,sha256=x1mWEpzSN2mc_HAUEaNEOUgS_Vh7Wa1VPwRFkRy4oG0,2574
 banks/filters/cache_control.py,sha256=aOGOIzuqasV_TcuFaaXbaoGhA2W9YTFuz7wkatyjXRU,962
-banks/filters/document.py,sha256=EhewpOgQ8ZW-vE1iaXQ-N0L6_xdWY-qHIp0VqL8wQDQ,2600
-banks/filters/image.py,sha256=Ls1fWCgRx0YLGIFx7hdKtR1skY575jDWlCESP0zV1Bs,1407
+banks/filters/document.py,sha256=hs2IO6d-xcLyTHH50bpzA3848bJLpQYjUkEgjOBYGqE,4893
+banks/filters/image.py,sha256=0t4u2El2Gi92C1qlY_0ji5OpPnjJfTn67SXb2mCIOl8,1507
 banks/filters/lemmatize.py,sha256=Yvp8M4HCx6C0nrcu3UEMtjJUwsyVYI6GQDYOG4S6EEw,887
 banks/filters/tool.py,sha256=i8ukSDYw54ksShVJ2abfRQAiKzKrqUtmgBB1H04cig0,475
-banks/filters/video.py,sha256=mD4iLWfFeihccqhFIr3Vq2_3ymF1rLMMuYzvTHEBlgk,2024
+banks/filters/video.py,sha256=MFni5um9Xnq8Sxf6ZBTN5GsKAC6f73CLrqFAGaE2pkk,3531
 banks/filters/xml.py,sha256=uQ_2zfCf8NhpdbF8F5HS7URXvDzsxfg-TEIVGufZbM0,1991
 banks/registries/__init__.py,sha256=iRK-8420cKBckOTd5KcIFQyV66EsF0Mc7UHCkzf8qZU,255
 banks/registries/directory.py,sha256=gRFO7fl9yXHt2NJ1pDA2wPSQtlORhSw1GKWxSTyFzE8,6055
 banks/registries/file.py,sha256=8ayvFrcM8Tk0DWgGXmKD2DRBfGXr5CmgtdQaQ5cXhow,4054
 banks/registries/redis.py,sha256=eBL92URJa-NegOxRLS4b2xrDRDxz6iiaz_7Ddi32Rtc,2756
-banks-2.3.0.dist-info/METADATA,sha256=-zwz3vftKvn3mKTVlICOKu7u4JEuyiEf8idRQf1-BEg,12227
-banks-2.3.0.dist-info/WHEEL,sha256=WLgqFyCfm_KASv4WHyYy0P3pM_m7J5L9k2skdKLirC8,87
-banks-2.3.0.dist-info/licenses/LICENSE.txt,sha256=NZJne_JTwMFwq_g-kq-sm4PuaeVOgu1l3NUGOgBHX-g,1102
-banks-2.3.0.dist-info/RECORD,,
+banks-2.4.0.dist-info/METADATA,sha256=D94dNSjkJj9vXMecMp7lTiGS_7sM4ipxyRvtbkn8aYs,12258
+banks-2.4.0.dist-info/WHEEL,sha256=WLgqFyCfm_KASv4WHyYy0P3pM_m7J5L9k2skdKLirC8,87
+banks-2.4.0.dist-info/licenses/LICENSE.txt,sha256=NZJne_JTwMFwq_g-kq-sm4PuaeVOgu1l3NUGOgBHX-g,1102
+banks-2.4.0.dist-info/RECORD,,