Add VHL audit file upload UI

app.py

@@ -6,14 +6,73 @@ import gradio as gr
 import requests
 
 HF_ROUTER_URL = "https://router.huggingface.co/v1/chat/completions"
+ROUTER_TIMEOUT_SECONDS = 600
+MAX_UPLOADED_FILES = 6
+MAX_UPLOADED_FILE_CHARS = 1_200_000
 DEFAULT_MODEL = "zai-org/GLM-5.1:zai-org"
 MODEL_CHOICES = [
     "zai-org/GLM-5.1:zai-org",
     "zai-org/GLM-5.1:fireworks-ai",
     "zai-org/GLM-5.1:novita",
 ]
-DEFAULT_SYSTEM = "You are a strict JSON extraction engine. Return the final answer only."
-DEFAULT_USER = "What is the capital of France?"
+DEFAULT_SYSTEM = """You are a strict clinical phenotype extraction auditor.
+Return valid JSON only. Do not rewrite the full extraction unless explicitly asked."""
+DEFAULT_USER = """Audit the uploaded grounded disease-layer output against the uploaded source sentence index.
+
+Check:
+- all evidence ids resolve to source sentences
+- no phenotype row aggregates unrelated sentences
+- qualifiers are supported by the exact evidence sentence
+- lab, imaging, test, management, inheritance, gene, prevalence, prognosis, and therapy items are routed outside phenotypes
+- ambiguous standalone targets are not left contextless
+- trigger factors are specific and point to a defined episode or phenotype context
+- ancillary and context rows stay grounded
+- missing clinically important findings from the source are listed only when clearly supported
+
+Return this JSON shape:
+{
+  "overall": "pass|minor_issues|major_issues",
+  "high_priority_fixes": [],
+  "missing_supported_findings": [],
+  "schema_or_grounding_violations": [],
+  "rows_to_downgrade_or_omit": [],
+  "notes": []
+}"""
+
+
+def _uploaded_file_path(uploaded_file: Any) -> str:
+    if hasattr(uploaded_file, "name"):
+        return str(uploaded_file.name)
+    return str(uploaded_file)
+
+
+def _read_uploaded_files(uploaded_files: Any) -> str:
+    if not uploaded_files:
+        return ""
+
+    files = uploaded_files if isinstance(uploaded_files, list) else [uploaded_files]
+    chunks = []
+    for uploaded_file in files[:MAX_UPLOADED_FILES]:
+        path = _uploaded_file_path(uploaded_file)
+        filename = os.path.basename(path)
+        try:
+            with open(path, "r", encoding="utf-8", errors="replace") as handle:
+                text = handle.read(MAX_UPLOADED_FILE_CHARS + 1)
+        except OSError as exc:
+            chunks.append(
+                f'\n\n<uploaded_file name="{filename}" read_error="{type(exc).__name__}">'
+                f"\n{exc}\n</uploaded_file>"
+            )
+            continue
+
+        truncated = len(text) > MAX_UPLOADED_FILE_CHARS
+        if truncated:
+            text = text[:MAX_UPLOADED_FILE_CHARS]
+        chunks.append(
+            f'\n\n<uploaded_file name="{filename}" truncated="{str(truncated).lower()}">'
+            f"\n{text}\n</uploaded_file>"
+        )
+    return "".join(chunks)
 
 
 def _extract_text(response_json: dict[str, Any]) -> tuple[str, str]:
@@ -30,7 +89,15 @@ def _extract_text(response_json: dict[str, Any]) -> tuple[str, str]:
     return str(content), str(reasoning)
 
 
-def call_router(model: str, system_prompt: str, user_prompt: str, max_completion_tokens: int, temperature: float, json_mode: bool):
+def call_router(
+    model: str,
+    system_prompt: str,
+    user_prompt: str,
+    uploaded_files: Any,
+    max_completion_tokens: int,
+    temperature: float,
+    json_mode: bool,
+):
     token = os.getenv("HF_TOKEN")
     if not token:
         return (
@@ -45,6 +112,10 @@ def call_router(model: str, system_prompt: str, user_prompt: str, max_completion
             ),
         )
 
+    file_context = _read_uploaded_files(uploaded_files)
+    if file_context:
+        user_prompt = f"{user_prompt.rstrip()}\n\n<uploaded_files>{file_context}\n</uploaded_files>"
+
     body: dict[str, Any] = {
         "model": model,
         "stream": False,
@@ -65,7 +136,7 @@ def call_router(model: str, system_prompt: str, user_prompt: str, max_completion
             "Content-Type": "application/json",
         },
         json=body,
-        timeout=180,
+        timeout=ROUTER_TIMEOUT_SECONDS,
     )
 
     try:
@@ -83,7 +154,7 @@ def call_router(model: str, system_prompt: str, user_prompt: str, max_completion
 
 with gr.Blocks(title="GLM-5.1 Lab") as demo:
     gr.Markdown("# GLM-5.1 Lab")
-    gr.Markdown("Direct Hugging Face router probe for `GLM-5.1` provider variants.")
+    gr.Markdown("Direct Hugging Face router probe for `GLM-5.1` provider variants. Upload JSON files to append them to the user prompt.")
 
     with gr.Row():
         model = gr.Dropdown(MODEL_CHOICES, value=DEFAULT_MODEL, label="Model")
@@ -93,6 +164,7 @@ with gr.Blocks(title="GLM-5.1 Lab") as demo:
 
     system_prompt = gr.Textbox(value=DEFAULT_SYSTEM, lines=4, label="System prompt")
     user_prompt = gr.Textbox(value=DEFAULT_USER, lines=12, label="User prompt")
+    uploaded_files = gr.File(file_count="multiple", type="filepath", label="Optional files to append")
     run_btn = gr.Button("Run")
 
     content_out = gr.Textbox(lines=10, label="Assistant content")
@@ -101,7 +173,7 @@ with gr.Blocks(title="GLM-5.1 Lab") as demo:
 
     run_btn.click(
         fn=call_router,
-        inputs=[model, system_prompt, user_prompt, max_tokens, temperature, json_mode],
+        inputs=[model, system_prompt, user_prompt, uploaded_files, max_tokens, temperature, json_mode],
         outputs=[content_out, reasoning_out, raw_out],
     )