feat(cli): add --document-dpi and --document-max-pages to vlmrun gw chat

tests/test_gateway.py

@@ -1025,6 +1025,84 @@ def test_chat_extra_routes_gateway_field_to_extra_body(
         assert call["extra_body"] == {"document_dpi": 200}
         assert call["temperature"] == 0
 
+    def test_chat_document_dpi_rides_in_extra_body(self, runner, patched_cli, tmp_path):
+        pdf = tmp_path / "doc.pdf"
+        pdf.write_bytes(b"%PDF fake")
+        result = runner.invoke(
+            app,
+            ["gw", "chat", str(pdf), "-m", "pp-ocrv6", "--document-dpi", "150", "-ns"],
+        )
+        assert result.exit_code == 0, result.stdout
+        call = patched_cli["client"].gateway.completions.calls[-1]
+        assert call["extra_body"] == {"document_dpi": 150}
+        # A gateway field must not leak into create()'s own kwargs.
+        assert "document_dpi" not in call
+
+    def test_chat_document_dpi_short_flag(self, runner, patched_cli, tmp_path):
+        pdf = tmp_path / "doc.pdf"
+        pdf.write_bytes(b"%PDF fake")
+        result = runner.invoke(
+            app, ["gw", "chat", str(pdf), "-m", "pp-ocrv6", "-d", "300", "-ns"]
+        )
+        assert result.exit_code == 0, result.stdout
+        assert patched_cli["client"].gateway.completions.calls[-1]["extra_body"] == {
+            "document_dpi": 300
+        }
+
+    def test_chat_document_max_pages_rides_in_extra_body(
+        self, runner, patched_cli, tmp_path
+    ):
+        pdf = tmp_path / "doc.pdf"
+        pdf.write_bytes(b"%PDF fake")
+        result = runner.invoke(
+            app,
+            [
+                "gw",
+                "chat",
+                str(pdf),
+                "-m",
+                "pp-ocrv6",
+                "--document-max-pages",
+                "10",
+                "-ns",
+            ],
+        )
+        assert result.exit_code == 0, result.stdout
+        assert patched_cli["client"].gateway.completions.calls[-1]["extra_body"] == {
+            "document_max_pages": 10
+        }
+
+    def test_chat_dpi_merges_with_method_and_extra(self, runner, patched_cli, tmp_path):
+        pdf = tmp_path / "doc.pdf"
+        pdf.write_bytes(b"%PDF fake")
+        result = runner.invoke(
+            app,
+            [
+                "gw",
+                "chat",
+                str(pdf),
+                "-m",
+                "pp-ocrv6",
+                "--method",
+                "ocr",
+                "--document-dpi",
+                "150",
+                "--document-max-pages",
+                "50",
+                "-e",
+                "temperature=0",
+                "-ns",
+            ],
+        )
+        assert result.exit_code == 0, result.stdout
+        call = patched_cli["client"].gateway.completions.calls[-1]
+        assert call["extra_body"] == {
+            "method": "ocr",
+            "document_dpi": 150,
+            "document_max_pages": 50,
+        }
+        assert call["temperature"] == 0
+
     def test_chat_no_extra_body_when_unused(self, runner, patched_cli, tmp_path):
         f = tmp_path / "img.png"
         f.write_bytes(b"fakepng")

vlmrun/cli/_cli/gateway.py

@@ -12,6 +12,11 @@
 such as ``zai-org/glm-ocr`` and ``paddleocr/pp-ocrv6``) do not accept
 text-only input.
 
+Multi-page PDFs are sent whole; the gateway rasterizes and fans them out per
+page. ``--document-dpi`` tunes that rasterization (default 72; raise to
+150-300 for dense or high-resolution pages) and ``--document-max-pages`` caps
+how many pages are processed.
+
 Commands: ``health``, ``models`` (list or detail one model), ``chat``,
 ``embed`` (embeddings) and ``transcribe`` (audio transcriptions).
 """
@@ -55,6 +60,7 @@
   vlmrun gw chat a.pdf b.pdf -m paddleocr/pp-ocrv6
   vlmrun gw chat img.jpg -m paddleocr/pp-ocrv6
   vlmrun gw chat img.jpg -p "describe this image" -m qwen/qwen3.5-0.8b
+  vlmrun gw chat doc.pdf -m paddleocr/pp-ocrv6 --document-dpi 150
   vlmrun gw chat doc.pdf -m zai-org/glm-ocr -e temperature=0 -e max_tokens=4096
 
 \b
@@ -71,6 +77,8 @@
   aliases (e.g. `glm-ocr`) also work.
   Most gateway models (e.g. OCR models) require at least one input file and do
   not accept text-only prompts. Use -p only for models that support it.
+  PDFs are sent whole; the gateway rasterizes and fans them out per page. Raise
+  `--document-dpi` (default 72) to 150-300 if small or dense text is missed.
 """
 
 GATEWAY_HELP = """OCR, VLM, embedding and transcription models on the VLM Run gateway.
@@ -562,6 +570,21 @@ def chat(
         "--method-params",
         help='JSON object of method arguments, e.g. \'{"lang": "en"}\'.',
     ),
+    document_dpi: Optional[int] = typer.Option(
+        None,
+        "--document-dpi",
+        "-d",
+        help=(
+            "Rasterization DPI per PDF page (gateway default 72; 150 is a good "
+            "balance; 300+ preserves fine print on dense/high-res pages at "
+            "higher cost). PDF input only."
+        ),
+    ),
+    document_max_pages: Optional[int] = typer.Option(
+        None,
+        "--document-max-pages",
+        help="Cap the number of PDF pages processed (gateway default 500).",
+    ),
     response_format: Optional[str] = typer.Option(
         None,
         "--response-format",
@@ -614,6 +637,13 @@ def chat(
             raise typer.Exit(1)
         extra_body["method_params"] = parsed_params
 
+    # Document rasterization controls are gateway-specific, so they ride in
+    # extra_body as top-level request-body fields (like `method`).
+    if document_dpi is not None:
+        extra_body["document_dpi"] = document_dpi
+    if document_max_pages is not None:
+        extra_body["document_max_pages"] = document_max_pages
+
     if response_format:
         # A standard OpenAI create() field, so it rides as a top-level kwarg.
         create_kwargs["response_format"] = _parse_response_format(response_format)