Update inline math line merges

marker/builders/line.py

@@ -113,7 +113,11 @@ class LineBuilder(BaseBuilder):
     def __call__(self, document: Document, provider: PdfProvider):
         # Disable Inline Detection for documents where layout model doesn't detect any equations
         # Also disable if we won't use the inline detections (if we aren't using the LLM or texify)
-        do_inline_math_detection = document.contained_blocks([BlockTypes.Equation, BlockTypes.TextInlineMath]) and (self.texify_inline_spans or self.use_llm)
+        do_inline_math_detection = all([
+            len(document.contained_blocks([BlockTypes.Equation, BlockTypes.TextInlineMath])) > 0,
+            (self.texify_inline_spans or self.use_llm)
+        ])
+
         provider_lines, ocr_lines = self.get_all_lines(document, provider, do_inline_math_detection)
         self.merge_blocks(document, provider_lines, ocr_lines)
 
@@ -186,7 +190,7 @@ class LineBuilder(BaseBuilder):
         if sum(layout_good) > len(document.pages) * self.min_document_ocr_threshold:
             layout_good = [True] * len(document.pages)
 
-        run_detection = [not good or do_inline_math_detection for good in layout_good]
+        run_detection = [(not good or do_inline_math_detection) for good in layout_good]
         page_images = [page.get_image(highres=False, remove_blocks=self.ocr_remove_blocks) for page, good in zip(document.pages, run_detection) if good]
 
         # Note: run_detection is longer than page_images, since it has a value for each page, not just good ones
@@ -416,13 +420,29 @@ class LineBuilder(BaseBuilder):
             best_overlap = np.argmax(overlaps[i])
             merge_lines[best_overlap].append(i)
 
+        # Filter to get rid of detected lines that include multiple provider lines
+        filtered_merge_lines = {}
+        for line_idx in merge_lines:
+            first_line = horizontal_provider_lines[merge_lines[line_idx][0]][1].line.polygon
+            all_close = all([
+                (
+                    abs(horizontal_provider_lines[ml][1].line.polygon.y_start - first_line.y_start) < self.inline_math_line_vertical_merge_threshold
+                    or
+                    abs(horizontal_provider_lines[ml][1].line.polygon.y_end - first_line.y_end) < self.inline_math_line_vertical_merge_threshold
+                )
+                for ml in
+                merge_lines[line_idx]
+            ])
+            if all_close:
+                filtered_merge_lines[line_idx] = merge_lines[line_idx]
+
         # Handle the merging
         already_merged = set()
-        potential_merges = set(chain.from_iterable(merge_lines.values()))
+        potential_merges = set(chain.from_iterable(filtered_merge_lines.values()))
         out_provider_lines = [(i, p) for i, p in enumerate(provider_lines) if i not in potential_merges]
-        for line_idx in merge_lines:
+        for line_idx in filtered_merge_lines:
             text_line = text_lines[line_idx]
-            merge_section = merge_lines[line_idx]
+            merge_section = filtered_merge_lines[line_idx]
             merge_section = [m for m in merge_section if m not in already_merged]
             if len(merge_section) == 0:
                 continue

marker/processors/debug.py

@@ -72,15 +72,20 @@ class DebugProcessor(BaseProcessor):
 
             line_bboxes = []
             span_bboxes = []
+            line_ids = []
             for child in page.children:
+                # Skip any blocks that have been removed
+                if child.removed:
+                    continue
                 if child.block_type == BlockTypes.Line:
                     bbox = child.polygon.rescale(page.polygon.size, png_image.size).bbox
                     line_bboxes.append(bbox)
+                    line_ids.append(child.block_id)
                 elif child.block_type == BlockTypes.Span:
                     bbox = child.polygon.rescale(page.polygon.size, png_image.size).bbox
                     span_bboxes.append(bbox)
 
-            self.render_on_image(line_bboxes, png_image, color="blue", draw_bbox=True, label_font_size=24)
+            self.render_on_image(line_bboxes, png_image, color="blue", draw_bbox=True, label_font_size=24, labels=[str(i) for i in line_ids])
             #self.render_on_image(span_bboxes, png_image, color="green", draw_bbox=True, label_font_size=24)
 
             png_image = self.render_layout_boxes(page, png_image)

marker/processors/line_merge.py

@@ -23,7 +23,7 @@ class LineMergeProcessor(BaseProcessor):
     intersection_pct_threshold: Annotated[
         float,
         "The total amount of intersection area concentrated in the max intersection block."
-    ] = .9
+    ] = .7
 
     def __init__(self, config):
         super().__init__(config)
@@ -38,7 +38,8 @@ class LineMergeProcessor(BaseProcessor):
                     continue
 
                 lines = block.contained_blocks(document, (BlockTypes.Line,))
-                line_bboxes = [l.polygon.bbox for l in lines]
+                lines = [l for l in lines if l.polygon.width * 5 > l.polygon.height]  # Skip vertical lines
+                line_bboxes = [l.polygon.expand(self.min_merge_pct, 0).bbox for l in lines] # Expand horizontally
                 intersections = matrix_intersection_area(line_bboxes, line_bboxes)
 
                 merges = []
@@ -49,6 +50,10 @@ class LineMergeProcessor(BaseProcessor):
                     intersection_pct = intersection_val / max(1, lines[i].polygon.area)
                     intersection_row = intersections[i]
                     intersection_row[i] = 0 # Zero out the current idx
+
+                    # Zero out previous merge segments
+                    for m in merge:
+                        intersection_row[m] = 0
                     max_intersection_idx = intersection_row.argmax()
                     total_intersection = max(1, sum(intersection_row))
                     max_intersection = intersection_row[max_intersection_idx]
@@ -61,7 +66,9 @@ class LineMergeProcessor(BaseProcessor):
                         abs(lines[i].polygon.y_end - lines[next_idx].polygon.y_end) <= self.min_merge_ydist,
                         max_intersection / total_intersection >= self.intersection_pct_threshold
                     ]):
-                        merge.append(i)
+                        if not merge:
+                            merge.append(i)
+                        merge.append(next_idx)
                     else:
                         merges.append(merge)
                         merge = []
@@ -82,6 +89,7 @@ class LineMergeProcessor(BaseProcessor):
                         other_line: Line = lines[idx]
                         line.merge(other_line)
                         block.structure.remove(other_line.id)
+                        other_line.removed = True # Mark line as removed
                         merged.add(idx)
 
                     # It is probably math if we are merging provider lines like this

marker/processors/llm/llm_inlinemath.py

@@ -83,11 +83,14 @@ Output:
         if not self.redo_inline_math:
             return
 
+        # Get inline math blocks
         inline_blocks = [
             (page, block)
             for page in document.pages
             for block in page.contained_blocks(document, self.block_types)
         ]
+
+        # Get other blocks with detected math in them
         detected_blocks = [
             (page, block)
             for page in document.pages
@@ -95,6 +98,7 @@ Output:
             if any([b.formats and "math" in b.formats for b in block.contained_blocks(document, (BlockTypes.Line,))])
         ]
         inference_blocks = inline_blocks + detected_blocks
+
         # Don't show progress if there are no blocks to process
         total_blocks = len(inference_blocks)
         if total_blocks == 0:

marker/schema/blocks/base.py

@@ -215,7 +215,7 @@ class Block(BaseModel):
         blocks = []
         for block_id in self.structure:
             block = document.get_block(block_id)
-            if block_types is None or block.block_type in block_types:
+            if (block_types is None or block.block_type in block_types) and not block.removed:
                 blocks.append(block)
             blocks += block.contained_blocks(document, block_types)
         return blocks

marker/services/claude.py

@@ -0,0 +1,133 @@
+import base64
+import json
+import time
+from io import BytesIO
+from typing import List, Annotated, Union, T
+
+import PIL
+from PIL import Image
+import anthropic
+from anthropic import APIError, APIConnectionError, APITimeoutError, RateLimitError
+from pydantic import BaseModel
+
+from marker.schema.blocks import Block
+from marker.services import BaseService
+
+class ClaudeService(BaseService):
+    claude_model_name: Annotated[
+        str,
+        "The name of the Google model to use for the service."
+    ] = "claude-3-5-sonnet-20241022"
+    claude_api_key: Annotated[
+        str,
+        "The Claude API key to use for the service."
+    ] = None
+    max_claude_tokens: Annotated[
+        int,
+        "The maximum number of tokens to use for a single Claude request."
+    ] = 4096
+
+
+    def img_to_base64(self, img: PIL.Image.Image):
+        image_bytes = BytesIO()
+        img.save(image_bytes, format="WEBP")
+        return base64.b64encode(image_bytes.getvalue()).decode('utf-8')
+
+    def prepare_images(self, images: Union[Image.Image, List[Image.Image]]) -> List[dict]:
+        if isinstance(images, Image.Image):
+            images = [images]
+
+        return [
+            {
+                "type": "image",
+                "source": {
+                    "type": "base64",
+                    "media_type": "image/webp",
+                    "data": self.img_to_base64(img)
+                }
+            }
+            for img in images
+        ]
+
+    def validate_response(self, response_text: str, schema: type[T]) -> T:
+        try:
+            # Try to parse as JSON first
+            data = json.loads(response_text)
+            return schema.parse_obj(data)
+        except json.JSONDecodeError:
+            # If not JSON, try to parse the raw text into the schema
+            return schema.parse_raw(response_text)
+
+    def get_client(self):
+        return anthropic.Anthropic(
+            api_key=self.claude_api_key,
+        )
+
+    def __call__(
+            self,
+            prompt: str,
+            image: PIL.Image.Image | List[PIL.Image.Image],
+            block: Block,
+            response_schema: type[BaseModel],
+            max_retries: int | None = None,
+            timeout: int | None = None
+    ):
+        if max_retries is None:
+            max_retries = self.max_retries
+
+        if timeout is None:
+            timeout = self.timeout
+
+        if not isinstance(image, list):
+            image = [image]
+
+        schema_example = response_schema.model_json_schema()
+        system_prompt = f"""
+Follow the instructions given by the user prompt.  You must provide your response in JSON format matching this schema:
+
+{json.dumps(schema_example, indent=2)}
+""".strip()
+
+        client = self.get_client()
+        image_data = self.prepare_images(image)
+
+        messages = [
+            {
+                "role": "system",
+                "content": system_prompt
+            },
+            {
+                "role": "user",
+                "content": [
+                    {
+                        "type": "text",
+                        "text": prompt
+                    },
+                    *image_data
+                ]
+            }
+        ]
+
+        tries = 0
+        while tries < max_retries:
+            try:
+                response = client.messages.create(
+                    model=self.claude_model_name,
+                    max_tokens=self.max_claude_tokens,
+                    messages=messages,
+                    timeout=timeout
+                )
+                # Extract and validate response
+                response_text = response.content[0].text
+                return self.validate_response(response_text, response_schema)
+            except RateLimitError as e:
+                # Rate limit exceeded
+                tries += 1
+                wait_time = tries * 3
+                print(f"Rate limit error: {e}. Retrying in {wait_time} seconds... (Attempt {tries}/{max_retries})")
+                time.sleep(wait_time)
+            except Exception as e:
+                print(e)
+                break
+
+        return {}

poetry.lock

@@ -163,6 +163,30 @@ files = [
     {file = "annotated_types-0.7.0.tar.gz", hash = "sha256:aff07c09a53a08bc8cfccb9c85b05f1aa9a2a6f23728d790723543408344ce89"},
 ]
 
+[[package]]
+name = "anthropic"
+version = "0.46.0"
+description = "The official Python library for the anthropic API"
+optional = false
+python-versions = ">=3.8"
+files = [
+    {file = "anthropic-0.46.0-py3-none-any.whl", hash = "sha256:1445ec9be78d2de7ea51b4d5acd3574e414aea97ef903d0ecbb57bec806aaa49"},
+    {file = "anthropic-0.46.0.tar.gz", hash = "sha256:eac3d43271d02321a57c3ca68aca84c3d58873e8e72d1433288adee2d46b745b"},
+]
+
+[package.dependencies]
+anyio = ">=3.5.0,<5"
+distro = ">=1.7.0,<2"
+httpx = ">=0.23.0,<1"
+jiter = ">=0.4.0,<1"
+pydantic = ">=1.9.0,<3"
+sniffio = "*"
+typing-extensions = ">=4.10,<5"
+
+[package.extras]
+bedrock = ["boto3 (>=1.28.57)", "botocore (>=1.31.57)"]
+vertex = ["google-auth (>=2,<3)"]
+
 [[package]]
 name = "anyio"
 version = "4.8.0"
@@ -777,6 +801,17 @@ files = [
     {file = "Distance-0.1.3.tar.gz", hash = "sha256:60807584f5b6003f5c521aa73f39f51f631de3be5cccc5a1d67166fcbf0d4551"},
 ]
 
+[[package]]
+name = "distro"
+version = "1.9.0"
+description = "Distro - an OS platform information API"
+optional = false
+python-versions = ">=3.6"
+files = [
+    {file = "distro-1.9.0-py3-none-any.whl", hash = "sha256:7bffd925d65168f85027d8da9af6bddab658135b840670a223589bc0c8ef02b2"},
+    {file = "distro-1.9.0.tar.gz", hash = "sha256:2fa77c6fd8940f116ee1d6b94a2f90b13b5ea8d019b98bc8bafdcabcdd9bdbed"},
+]
+
 [[package]]
 name = "exceptiongroup"
 version = "1.2.2"
@@ -1450,6 +1485,91 @@ MarkupSafe = ">=2.0"
 [package.extras]
 i18n = ["Babel (>=2.7)"]
 
+[[package]]
+name = "jiter"
+version = "0.8.2"
+description = "Fast iterable JSON parser."
+optional = false
+python-versions = ">=3.8"
+files = [
+    {file = "jiter-0.8.2-cp310-cp310-macosx_10_12_x86_64.whl", hash = "sha256:ca8577f6a413abe29b079bc30f907894d7eb07a865c4df69475e868d73e71c7b"},
+    {file = "jiter-0.8.2-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:b25bd626bde7fb51534190c7e3cb97cee89ee76b76d7585580e22f34f5e3f393"},
+    {file = "jiter-0.8.2-cp310-cp310-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:d5c826a221851a8dc028eb6d7d6429ba03184fa3c7e83ae01cd6d3bd1d4bd17d"},
+    {file = "jiter-0.8.2-cp310-cp310-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:d35c864c2dff13dfd79fb070fc4fc6235d7b9b359efe340e1261deb21b9fcb66"},
+    {file = "jiter-0.8.2-cp310-cp310-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:f557c55bc2b7676e74d39d19bcb8775ca295c7a028246175d6a8b431e70835e5"},
+    {file = "jiter-0.8.2-cp310-cp310-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:580ccf358539153db147e40751a0b41688a5ceb275e6f3e93d91c9467f42b2e3"},
+    {file = "jiter-0.8.2-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:af102d3372e917cffce49b521e4c32c497515119dc7bd8a75665e90a718bbf08"},
+    {file = "jiter-0.8.2-cp310-cp310-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:cadcc978f82397d515bb2683fc0d50103acff2a180552654bb92d6045dec2c49"},
+    {file = "jiter-0.8.2-cp310-cp310-musllinux_1_1_aarch64.whl", hash = "sha256:ba5bdf56969cad2019d4e8ffd3f879b5fdc792624129741d3d83fc832fef8c7d"},
+    {file = "jiter-0.8.2-cp310-cp310-musllinux_1_1_x86_64.whl", hash = "sha256:3b94a33a241bee9e34b8481cdcaa3d5c2116f575e0226e421bed3f7a6ea71cff"},
+    {file = "jiter-0.8.2-cp310-cp310-win32.whl", hash = "sha256:6e5337bf454abddd91bd048ce0dca5134056fc99ca0205258766db35d0a2ea43"},
+    {file = "jiter-0.8.2-cp310-cp310-win_amd64.whl", hash = "sha256:4a9220497ca0cb1fe94e3f334f65b9b5102a0b8147646118f020d8ce1de70105"},
+    {file = "jiter-0.8.2-cp311-cp311-macosx_10_12_x86_64.whl", hash = "sha256:2dd61c5afc88a4fda7d8b2cf03ae5947c6ac7516d32b7a15bf4b49569a5c076b"},
+    {file = "jiter-0.8.2-cp311-cp311-macosx_11_0_arm64.whl", hash = "sha256:a6c710d657c8d1d2adbbb5c0b0c6bfcec28fd35bd6b5f016395f9ac43e878a15"},
+    {file = "jiter-0.8.2-cp311-cp311-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:a9584de0cd306072635fe4b89742bf26feae858a0683b399ad0c2509011b9dc0"},
+    {file = "jiter-0.8.2-cp311-cp311-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:5a90a923338531b7970abb063cfc087eebae6ef8ec8139762007188f6bc69a9f"},
+    {file = "jiter-0.8.2-cp311-cp311-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:d21974d246ed0181558087cd9f76e84e8321091ebfb3a93d4c341479a736f099"},
+    {file = "jiter-0.8.2-cp311-cp311-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:32475a42b2ea7b344069dc1e81445cfc00b9d0e3ca837f0523072432332e9f74"},
+    {file = "jiter-0.8.2-cp311-cp311-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:8b9931fd36ee513c26b5bf08c940b0ac875de175341cbdd4fa3be109f0492586"},
+    {file = "jiter-0.8.2-cp311-cp311-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:ce0820f4a3a59ddced7fce696d86a096d5cc48d32a4183483a17671a61edfddc"},
+    {file = "jiter-0.8.2-cp311-cp311-musllinux_1_1_aarch64.whl", hash = "sha256:8ffc86ae5e3e6a93765d49d1ab47b6075a9c978a2b3b80f0f32628f39caa0c88"},
+    {file = "jiter-0.8.2-cp311-cp311-musllinux_1_1_x86_64.whl", hash = "sha256:5127dc1abd809431172bc3fbe8168d6b90556a30bb10acd5ded41c3cfd6f43b6"},
+    {file = "jiter-0.8.2-cp311-cp311-win32.whl", hash = "sha256:66227a2c7b575720c1871c8800d3a0122bb8ee94edb43a5685aa9aceb2782d44"},
+    {file = "jiter-0.8.2-cp311-cp311-win_amd64.whl", hash = "sha256:cde031d8413842a1e7501e9129b8e676e62a657f8ec8166e18a70d94d4682855"},
+    {file = "jiter-0.8.2-cp312-cp312-macosx_10_12_x86_64.whl", hash = "sha256:e6ec2be506e7d6f9527dae9ff4b7f54e68ea44a0ef6b098256ddf895218a2f8f"},
+    {file = "jiter-0.8.2-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:76e324da7b5da060287c54f2fabd3db5f76468006c811831f051942bf68c9d44"},
+    {file = "jiter-0.8.2-cp312-cp312-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:180a8aea058f7535d1c84183c0362c710f4750bef66630c05f40c93c2b152a0f"},
+    {file = "jiter-0.8.2-cp312-cp312-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:025337859077b41548bdcbabe38698bcd93cfe10b06ff66617a48ff92c9aec60"},
+    {file = "jiter-0.8.2-cp312-cp312-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:ecff0dc14f409599bbcafa7e470c00b80f17abc14d1405d38ab02e4b42e55b57"},
+    {file = "jiter-0.8.2-cp312-cp312-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:ffd9fee7d0775ebaba131f7ca2e2d83839a62ad65e8e02fe2bd8fc975cedeb9e"},
+    {file = "jiter-0.8.2-cp312-cp312-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:14601dcac4889e0a1c75ccf6a0e4baf70dbc75041e51bcf8d0e9274519df6887"},
+    {file = "jiter-0.8.2-cp312-cp312-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:92249669925bc1c54fcd2ec73f70f2c1d6a817928480ee1c65af5f6b81cdf12d"},
+    {file = "jiter-0.8.2-cp312-cp312-musllinux_1_1_aarch64.whl", hash = "sha256:e725edd0929fa79f8349ab4ec7f81c714df51dc4e991539a578e5018fa4a7152"},
+    {file = "jiter-0.8.2-cp312-cp312-musllinux_1_1_x86_64.whl", hash = "sha256:bf55846c7b7a680eebaf9c3c48d630e1bf51bdf76c68a5f654b8524335b0ad29"},
+    {file = "jiter-0.8.2-cp312-cp312-win32.whl", hash = "sha256:7efe4853ecd3d6110301665a5178b9856be7e2a9485f49d91aa4d737ad2ae49e"},
+    {file = "jiter-0.8.2-cp312-cp312-win_amd64.whl", hash = "sha256:83c0efd80b29695058d0fd2fa8a556490dbce9804eac3e281f373bbc99045f6c"},
+    {file = "jiter-0.8.2-cp313-cp313-macosx_10_12_x86_64.whl", hash = "sha256:ca1f08b8e43dc3bd0594c992fb1fd2f7ce87f7bf0d44358198d6da8034afdf84"},
+    {file = "jiter-0.8.2-cp313-cp313-macosx_11_0_arm64.whl", hash = "sha256:5672a86d55416ccd214c778efccf3266b84f87b89063b582167d803246354be4"},
+    {file = "jiter-0.8.2-cp313-cp313-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:58dc9bc9767a1101f4e5e22db1b652161a225874d66f0e5cb8e2c7d1c438b587"},
+    {file = "jiter-0.8.2-cp313-cp313-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:37b2998606d6dadbb5ccda959a33d6a5e853252d921fec1792fc902351bb4e2c"},
+    {file = "jiter-0.8.2-cp313-cp313-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:4ab9a87f3784eb0e098f84a32670cfe4a79cb6512fd8f42ae3d0709f06405d18"},
+    {file = "jiter-0.8.2-cp313-cp313-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:79aec8172b9e3c6d05fd4b219d5de1ac616bd8da934107325a6c0d0e866a21b6"},
+    {file = "jiter-0.8.2-cp313-cp313-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:711e408732d4e9a0208008e5892c2966b485c783cd2d9a681f3eb147cf36c7ef"},
+    {file = "jiter-0.8.2-cp313-cp313-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:653cf462db4e8c41995e33d865965e79641ef45369d8a11f54cd30888b7e6ff1"},
+    {file = "jiter-0.8.2-cp313-cp313-musllinux_1_1_aarch64.whl", hash = "sha256:9c63eaef32b7bebac8ebebf4dabebdbc6769a09c127294db6babee38e9f405b9"},
+    {file = "jiter-0.8.2-cp313-cp313-musllinux_1_1_x86_64.whl", hash = "sha256:eb21aaa9a200d0a80dacc7a81038d2e476ffe473ffdd9c91eb745d623561de05"},
+    {file = "jiter-0.8.2-cp313-cp313-win32.whl", hash = "sha256:789361ed945d8d42850f919342a8665d2dc79e7e44ca1c97cc786966a21f627a"},
+    {file = "jiter-0.8.2-cp313-cp313-win_amd64.whl", hash = "sha256:ab7f43235d71e03b941c1630f4b6e3055d46b6cb8728a17663eaac9d8e83a865"},
+    {file = "jiter-0.8.2-cp313-cp313t-macosx_11_0_arm64.whl", hash = "sha256:b426f72cd77da3fec300ed3bc990895e2dd6b49e3bfe6c438592a3ba660e41ca"},
+    {file = "jiter-0.8.2-cp313-cp313t-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:b2dd880785088ff2ad21ffee205e58a8c1ddabc63612444ae41e5e4b321b39c0"},
+    {file = "jiter-0.8.2-cp313-cp313t-win_amd64.whl", hash = "sha256:3ac9f578c46f22405ff7f8b1f5848fb753cc4b8377fbec8470a7dc3997ca7566"},
+    {file = "jiter-0.8.2-cp38-cp38-macosx_10_12_x86_64.whl", hash = "sha256:9e1fa156ee9454642adb7e7234a383884452532bc9d53d5af2d18d98ada1d79c"},
+    {file = "jiter-0.8.2-cp38-cp38-macosx_11_0_arm64.whl", hash = "sha256:0cf5dfa9956d96ff2efb0f8e9c7d055904012c952539a774305aaaf3abdf3d6c"},
+    {file = "jiter-0.8.2-cp38-cp38-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:e52bf98c7e727dd44f7c4acb980cb988448faeafed8433c867888268899b298b"},
+    {file = "jiter-0.8.2-cp38-cp38-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:a2ecaa3c23e7a7cf86d00eda3390c232f4d533cd9ddea4b04f5d0644faf642c5"},
+    {file = "jiter-0.8.2-cp38-cp38-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:08d4c92bf480e19fc3f2717c9ce2aa31dceaa9163839a311424b6862252c943e"},
+    {file = "jiter-0.8.2-cp38-cp38-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:99d9a1eded738299ba8e106c6779ce5c3893cffa0e32e4485d680588adae6db8"},
+    {file = "jiter-0.8.2-cp38-cp38-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:d20be8b7f606df096e08b0b1b4a3c6f0515e8dac296881fe7461dfa0fb5ec817"},
+    {file = "jiter-0.8.2-cp38-cp38-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:d33f94615fcaf872f7fd8cd98ac3b429e435c77619777e8a449d9d27e01134d1"},
+    {file = "jiter-0.8.2-cp38-cp38-musllinux_1_1_aarch64.whl", hash = "sha256:317b25e98a35ffec5c67efe56a4e9970852632c810d35b34ecdd70cc0e47b3b6"},
+    {file = "jiter-0.8.2-cp38-cp38-musllinux_1_1_x86_64.whl", hash = "sha256:fc9043259ee430ecd71d178fccabd8c332a3bf1e81e50cae43cc2b28d19e4cb7"},
+    {file = "jiter-0.8.2-cp38-cp38-win32.whl", hash = "sha256:fc5adda618205bd4678b146612ce44c3cbfdee9697951f2c0ffdef1f26d72b63"},
+    {file = "jiter-0.8.2-cp38-cp38-win_amd64.whl", hash = "sha256:cd646c827b4f85ef4a78e4e58f4f5854fae0caf3db91b59f0d73731448a970c6"},
+    {file = "jiter-0.8.2-cp39-cp39-macosx_10_12_x86_64.whl", hash = "sha256:e41e75344acef3fc59ba4765df29f107f309ca9e8eace5baacabd9217e52a5ee"},
+    {file = "jiter-0.8.2-cp39-cp39-macosx_11_0_arm64.whl", hash = "sha256:7f22b16b35d5c1df9dfd58843ab2cd25e6bf15191f5a236bed177afade507bfc"},
+    {file = "jiter-0.8.2-cp39-cp39-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:f7200b8f7619d36aa51c803fd52020a2dfbea36ffec1b5e22cab11fd34d95a6d"},
+    {file = "jiter-0.8.2-cp39-cp39-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:70bf4c43652cc294040dbb62256c83c8718370c8b93dd93d934b9a7bf6c4f53c"},
+    {file = "jiter-0.8.2-cp39-cp39-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:f9d471356dc16f84ed48768b8ee79f29514295c7295cb41e1133ec0b2b8d637d"},
+    {file = "jiter-0.8.2-cp39-cp39-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:859e8eb3507894093d01929e12e267f83b1d5f6221099d3ec976f0c995cb6bd9"},
+    {file = "jiter-0.8.2-cp39-cp39-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:eaa58399c01db555346647a907b4ef6d4f584b123943be6ed5588c3f2359c9f4"},
+    {file = "jiter-0.8.2-cp39-cp39-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:8f2d5ed877f089862f4c7aacf3a542627c1496f972a34d0474ce85ee7d939c27"},
+    {file = "jiter-0.8.2-cp39-cp39-musllinux_1_1_aarch64.whl", hash = "sha256:03c9df035d4f8d647f8c210ddc2ae0728387275340668fb30d2421e17d9a0841"},
+    {file = "jiter-0.8.2-cp39-cp39-musllinux_1_1_x86_64.whl", hash = "sha256:8bd2a824d08d8977bb2794ea2682f898ad3d8837932e3a74937e93d62ecbb637"},
+    {file = "jiter-0.8.2-cp39-cp39-win32.whl", hash = "sha256:ca29b6371ebc40e496995c94b988a101b9fbbed48a51190a4461fcb0a68b4a36"},
+    {file = "jiter-0.8.2-cp39-cp39-win_amd64.whl", hash = "sha256:1c0dfbd1be3cbefc7510102370d86e35d1d53e5a93d48519688b1bf0f761160a"},
+    {file = "jiter-0.8.2.tar.gz", hash = "sha256:cd73d3e740666d0e639f678adb176fad25c1bcbdae88d8d7b857e1783bb4212d"},
+]
+
 [[package]]
 name = "joblib"
 version = "1.4.2"
@@ -2691,10 +2811,10 @@ files = [
 
 [package.dependencies]
 numpy = [
-    {version = ">=1.26.0", markers = "python_version >= \"3.12\""},
-    {version = ">=1.23.5", markers = "python_version >= \"3.11\" and python_version < \"3.12\""},
     {version = ">=1.21.4", markers = "python_version >= \"3.10\" and platform_system == \"Darwin\" and python_version < \"3.11\""},
     {version = ">=1.21.2", markers = "platform_system != \"Darwin\" and python_version >= \"3.10\" and python_version < \"3.11\""},
+    {version = ">=1.23.5", markers = "python_version >= \"3.11\" and python_version < \"3.12\""},
+    {version = ">=1.26.0", markers = "python_version >= \"3.12\""},
 ]
 
 [[package]]
@@ -2772,9 +2892,9 @@ files = [
 
 [package.dependencies]
 numpy = [
-    {version = ">=1.26.0", markers = "python_version >= \"3.12\""},
-    {version = ">=1.23.2", markers = "python_version == \"3.11\""},
     {version = ">=1.22.4", markers = "python_version < \"3.11\""},
+    {version = ">=1.23.2", markers = "python_version == \"3.11\""},
+    {version = ">=1.26.0", markers = "python_version >= \"3.12\""},
 ]
 python-dateutil = ">=2.8.2"
 pytz = ">=2020.1"
@@ -5453,4 +5573,4 @@ propcache = ">=0.2.0"
 [metadata]
 lock-version = "2.0"
 python-versions = "^3.10"
-content-hash = "04afa6e305c60db8d9f5d304d67f6c51e0415cee9b3c7d1171750f7dd787135b"
+content-hash = "664bbdbf1226d34671e66f96973588ea266c9e0aefaace7d67da5a7cd00f0eec"

pyproject.toml

@@ -35,6 +35,7 @@ markdown2 = "^2.5.2"
 filetype = "^1.2.0"
 scikit-learn = "^1.6.1"
 google-genai = "^1.0.0"
+anthropic = "^0.46.0"
 
 [tool.poetry.group.dev.dependencies]
 jupyter = "^1.0.0"