Feat: add VLM-boosted PDF parser (#6278)

### What problem does this PR solve?

Add VLM-boosted PDF parser if VLM is set.

### Type of change

- [x] New Feature (non-breaking change which adds functionality)

deepdoc/parser/figure_parser.py

@@ -0,0 +1,82 @@
+#
+#  Copyright 2025 The InfiniFlow Authors. All Rights Reserved.
+#
+#  Licensed under the Apache License, Version 2.0 (the "License");
+#  you may not use this file except in compliance with the License.
+#  You may obtain a copy of the License at
+#
+#      http://www.apache.org/licenses/LICENSE-2.0
+#
+#  Unless required by applicable law or agreed to in writing, software
+#  distributed under the License is distributed on an "AS IS" BASIS,
+#  WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+#  See the License for the specific language governing permissions and
+#  limitations under the License.
+#
+
+
+from rag.app.picture import vision_llm_chunk as picture_vision_llm_chunk
+from rag.prompts import vision_llm_figure_describe_prompt
+
+
+class VisionFigureParser:
+    def __init__(self, vision_model, figures_data, *args, **kwargs):
+        self.vision_model = vision_model
+        self._extract_figures_info(figures_data)
+        assert len(self.figures) == len(self.descriptions)
+        assert not self.positions or (len(self.figures) == len(self.positions))
+
+    def _extract_figures_info(self, figures_data):
+        self.figures = []
+        self.descriptions = []
+        self.positions = []
+
+        for item in figures_data:
+            # position
+            if len(item) == 2 and isinstance(item[1], list) and len(item[1]) == 1 and len(item[1][0]) == 5:
+                img_desc = item[0]
+                assert len(img_desc) == 2, "Should be (figure, [description])"
+                self.figures.append(img_desc[0])
+                self.descriptions.append(img_desc[1])
+                self.positions.append(item[1])
+            else:
+                assert len(item) == 2 and isinstance(item, tuple), f"get {len(item)=}, {item=}"
+                self.figures.append(item[0])
+                self.descriptions.append(item[1])
+
+    def _assemble(self):
+        self.assembled = []
+        self.has_positions = len(self.positions) != 0
+        for i in range(len(self.figures)):
+            figure = self.figures[i]
+            desc = self.descriptions[i]
+            pos = self.positions[i] if self.has_positions else None
+
+            figure_desc = (figure, desc)
+
+            if pos is not None:
+                self.assembled.append((figure_desc, pos))
+            else:
+                self.assembled.append((figure_desc,))
+
+        return self.assembled
+
+    def __call__(self, **kwargs):
+        callback = kwargs.get("callback", lambda prog, msg: None)
+
+        for idx, img_binary in enumerate(self.figures or []):
+            figure_num = idx  # 0-based
+
+            txt = picture_vision_llm_chunk(
+                binary=img_binary,
+                vision_model=self.vision_model,
+                prompt=vision_llm_figure_describe_prompt(),
+                callback=callback,
+            )
+
+            if txt:
+                self.descriptions[figure_num] = txt + "\n".join(self.descriptions[figure_num])
+
+        self._assemble()
+
+        return self.assembled

deepdoc/parser/pdf_parser.py

@@ -653,8 +653,7 @@ class RAGFlowPdfParser:
             b_["top"] = b["top"]
             self.boxes.pop(i)
 
-    def _extract_table_figure(self, need_image, ZM,
-                              return_html, need_position):
+    def _extract_table_figure(self, need_image, ZM, return_html, need_position, separate_tables_figures=False):
         tables = {}
         figures = {}
         # extract figure and table boxes
@@ -768,9 +767,6 @@ class RAGFlowPdfParser:
                     tk)
             self.boxes.pop(i)
 
-        res = []
-        positions = []
-
         def cropout(bxs, ltype, poss):
             nonlocal ZM
             pn = set([b["page_number"] - 1 for b in bxs])
@@ -818,6 +814,10 @@ class RAGFlowPdfParser:
                 height += img.size[1]
             return pic
 
+        res = []
+        positions = []
+        figure_results = []
+        figure_positions = []
         # crop figure out and add caption
         for k, bxs in figures.items():
             txt = "\n".join([b["text"] for b in bxs])
@@ -825,28 +825,46 @@ class RAGFlowPdfParser:
                 continue
 
             poss = []
-            res.append(
-                (cropout(
-                    bxs,
-                    "figure", poss),
-                 [txt]))
-            positions.append(poss)
+
+            if separate_tables_figures:
+                figure_results.append(
+                    (cropout(
+                        bxs,
+                        "figure", poss),
+                     [txt]))
+                figure_positions.append(poss)
+            else:
+                res.append(
+                    (cropout(
+                        bxs,
+                        "figure", poss),
+                     [txt]))
+                positions.append(poss)
 
         for k, bxs in tables.items():
             if not bxs:
                 continue
             bxs = Recognizer.sort_Y_firstly(bxs, np.mean(
                 [(b["bottom"] - b["top"]) / 2 for b in bxs]))
+
             poss = []
+
             res.append((cropout(bxs, "table", poss),
                         self.tbl_det.construct_table(bxs, html=return_html, is_english=self.is_english)))
             positions.append(poss)
 
-        assert len(positions) == len(res)
-
-        if need_position:
-            return list(zip(res, positions))
-        return res
+        if separate_tables_figures:
+            assert len(positions) + len(figure_positions) == len(res) + len(figure_results)
+            if need_position:
+                return list(zip(res, positions)), list(zip(figure_results, figure_positions))
+            else:
+                return res, figure_results
+        else:
+            assert len(positions) == len(res)
+            if need_position:
+                return list(zip(res, positions))
+            else:
+                return res
 
     def proj_match(self, line):
         if len(line) <= 2:

rag/app/naive.py

@@ -30,6 +30,7 @@ from api.db import LLMType
 from api.db.services.llm_service import LLMBundle
 from deepdoc.parser import DocxParser, ExcelParser, HtmlParser, JsonParser, MarkdownParser, PdfParser, TxtParser
 from deepdoc.parser.pdf_parser import PlainParser, VisionParser
+from deepdoc.parser.figure_parser import VisionFigureParser
 from rag.nlp import concat_img, find_codec, naive_merge, naive_merge_docx, rag_tokenizer, tokenize_chunks, tokenize_chunks_docx, tokenize_table
 from rag.utils import num_tokens_from_string
 
@@ -134,7 +135,7 @@ class Pdf(PdfParser):
         super().__init__()
 
     def __call__(self, filename, binary=None, from_page=0,
-                 to_page=100000, zoomin=3, callback=None):
+                 to_page=100000, zoomin=3, callback=None, separate_tables_figures=False):
         start = timer()
         first_start = start
         callback(msg="OCR started")
@@ -159,14 +160,19 @@ class Pdf(PdfParser):
         start = timer()
         self._text_merge()
         callback(0.67, "Text merged ({:.2f}s)".format(timer() - start))
-        tbls = self._extract_table_figure(True, zoomin, True, True)
-        # self._naive_vertical_merge()
-        self._concat_downward()
-        # self._filter_forpages()
 
-        logging.info("layouts cost: {}s".format(timer() - first_start))
-        return [(b["text"], self._line_tag(b, zoomin))
-                for b in self.boxes], tbls
+        if separate_tables_figures:
+            tbls, figures = self._extract_table_figure(True, zoomin, True, True, True)
+            self._concat_downward()
+            logging.info("layouts cost: {}s".format(timer() - first_start))
+            return [(b["text"], self._line_tag(b, zoomin)) for b in self.boxes], tbls, figures
+        else:
+            tbls = self._extract_table_figure(True, zoomin, True, True)
+            # self._naive_vertical_merge()
+            self._concat_downward()
+            # self._filter_forpages()
+            logging.info("layouts cost: {}s".format(timer() - first_start))
+            return [(b["text"], self._line_tag(b, zoomin)) for b in self.boxes], tbls
 
 
 class Markdown(MarkdownParser):
@@ -243,15 +249,32 @@ def chunk(filename, binary=None, from_page=0, to_page=100000,
 
         if layout_recognizer == "DeepDOC":
             pdf_parser = Pdf()
-        elif layout_recognizer == "Plain Text":
-            pdf_parser = PlainParser()
-        else:
-            vision_model = LLMBundle(kwargs["tenant_id"], LLMType.IMAGE2TEXT, llm_name=layout_recognizer, lang=lang)
-            pdf_parser = VisionParser(vision_model=vision_model, **kwargs)
 
-        sections, tables = pdf_parser(filename if not binary else binary, from_page=from_page, to_page=to_page,
-                                      callback=callback)
-        res = tokenize_table(tables, doc, is_english)
+            try:
+                vision_model = LLMBundle(kwargs["tenant_id"], LLMType.IMAGE2TEXT)
+            except Exception:
+                vision_model = None
+
+            if vision_model:
+                sections, tables, figures = pdf_parser(filename if not binary else binary, from_page=from_page, to_page=to_page, callback=callback, separate_tables_figures=True)
+                pdf_vision_parser = VisionFigureParser(vision_model=vision_model, figures_data=figures, **kwargs)
+                boosted_figures = pdf_vision_parser(callback=callback)
+                tables.extend(boosted_figures)
+            else:
+                sections, tables = pdf_parser(filename if not binary else binary, from_page=from_page, to_page=to_page, callback=callback)
+
+            res = tokenize_table(tables, doc, is_english)
+
+        else:
+            if layout_recognizer == "Plain Text":
+                pdf_parser = PlainParser()
+            else:
+                vision_model = LLMBundle(kwargs["tenant_id"], LLMType.IMAGE2TEXT, llm_name=layout_recognizer, lang=lang)
+                pdf_parser = VisionParser(vision_model=vision_model, **kwargs)
+
+            sections, tables = pdf_parser(filename if not binary else binary, from_page=from_page, to_page=to_page,
+                                          callback=callback)
+            res = tokenize_table(tables, doc, is_english)
 
     elif re.search(r"\.(csv|xlsx?)$", filename, re.IGNORECASE):
         callback(0.1, "Start to parse.")

rag/app/picture.py

@@ -86,4 +86,4 @@ def vision_llm_chunk(binary, vision_model, prompt=None, callback=None):
     except Exception as e:
         callback(-1, str(e))
 
-    return []
+    return ""

rag/prompts.py

@@ -393,3 +393,28 @@ FAILURE HANDLING:
 - If you do not detect valid content in the image, return an empty string.
 """
     return prompt_en
+
+
+def vision_llm_figure_describe_prompt() -> str:
+    prompt = """
+You are an expert visual data analyst. Analyze the image and provide a comprehensive description of its content. Focus on identifying the type of visual data representation (e.g., bar chart, pie chart, line graph, table, flowchart), its structure, and any text captions or labels included in the image.
+
+Tasks:
+1. Describe the overall structure of the visual representation. Specify if it is a chart, graph, table, or diagram.
+2. Identify and extract any axes, legends, titles, or labels present in the image. Provide the exact text where available.
+3. Extract the data points from the visual elements (e.g., bar heights, line graph coordinates, pie chart segments, table rows and columns).
+4. Analyze and explain any trends, comparisons, or patterns shown in the data.
+5. Capture any annotations, captions, or footnotes, and explain their relevance to the image.
+6. Only include details that are explicitly present in the image. If an element (e.g., axis, legend, or caption) does not exist or is not visible, do not mention it.
+
+Output format (include only sections relevant to the image content):
+- Visual Type: [Type]
+- Title: [Title text, if available]
+- Axes / Legends / Labels: [Details, if available]
+- Data Points: [Extracted data]
+- Trends / Insights: [Analysis and interpretation]
+- Captions / Annotations: [Text and relevance, if available]
+
+Ensure high accuracy, clarity, and completeness in your analysis, and includes only the information present in the image. Avoid unnecessary statements about missing elements.
+"""
+    return prompt