Deploy ClaimReady: on-device Gemma 3 via llama.cpp

.gitignore

@@ -0,0 +1,4 @@
+.env
+__pycache__/
+venv/
+*.pyc

README.md

@@ -1,15 +1,25 @@
 ---
-title: Claim Ready
-emoji: 🌍
-colorFrom: yellow
-colorTo: pink
+title: ClaimReady — Claim Submission Check
+emoji: 🏥
+colorFrom: blue
+colorTo: green
 sdk: gradio
-sdk_version: 6.18.0
-python_version: '3.12'
 app_file: app.py
 pinned: false
-license: apache-2.0
 short_description: Catch claim errors before you submit — runs locally
 ---
 
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+# 🏥 ClaimReady
+
+Pre-check hospital insurance **claim** and **pre-authorization** documents against the
+Standard Treatment Guidelines **before** they are uploaded to the government portal —
+catching compliance errors early to reduce rejections.
+
+**Runs fully on-device:** OCR + content verification are done by **Gemma 3 12B**
+(a small, ≤32B open model) through **llama.cpp** — no cloud APIs.
+
+## How it works
+1. Pick a package code and stage (pre-auth / claim).
+2. Upload the supporting documents (images or PDFs).
+3. The model OCRs each file and checks it against the required-document checklist
+   and the STG content rules, returning a ✅ / ❌ / ⚠️ verdict.

app.py

@@ -0,0 +1,164 @@
+"""
+Hospital Claim Verification — PMJAY compliance pre-check.
+
+A Gradio app for hospital billing staff to verify PMJAY pre-auth / claim document
+sets against the Standard Treatment Guidelines BEFORE uploading to the Government
+of India portal — catching compliance errors early to reduce rejections.
+
+Step 2: verification wired to Gemma 3 12B (via OpenRouter) for OCR + content
+checks against the Standard Treatment Guidelines.
+"""
+
+import gradio as gr
+
+from llm import LLMConfigError, verify
+from packages import (
+    STAGE_PREAUTH,
+    STAGE_CLAIM,
+    STAGE_LABELS,
+    dropdown_choices,
+    get_package,
+    required_documents,
+)
+
+STAGE_RADIO_CHOICES = [
+    (STAGE_LABELS[STAGE_PREAUTH], STAGE_PREAUTH),
+    (STAGE_LABELS[STAGE_CLAIM], STAGE_CLAIM),
+]
+
+
+def package_info_md(code):
+    """Markdown summary of the selected package."""
+    pkg = get_package(code)
+    if not pkg:
+        return ""
+    return (
+        f"**{pkg['code']} — {pkg['name']}**  \n"
+        f"Procedure: {pkg.get('procedure', '—')}  \n"
+        f"Specialty: {pkg.get('specialty', '—')}  \n"
+        f"Package price: {pkg.get('price', '—')}"
+    )
+
+
+def checklist_md(code, stage):
+    """Markdown checklist of required documents for a package + stage."""
+    if not code or not stage:
+        return "_Select a package and stage to see the required documents._"
+    pkg = get_package(code)
+    docs = required_documents(code, stage)
+    if not docs:
+        return "_No documents configured for this selection._"
+
+    lines = [
+        f"### Required documents — {STAGE_LABELS[stage]}",
+        f"_{pkg['name']} ({pkg['code']})_",
+        "",
+    ]
+    for i, doc in enumerate(docs, 1):
+        line = f"{i}. **{doc['label']}** — {doc['verify']}"
+        if doc.get("note"):
+            line += f"  \n   _Note: {doc['note']}_"
+        lines.append(line)
+
+    rules = pkg.get("rules", [])
+    if rules:
+        lines.append("")
+        lines.append("**Compliance checks (content):**")
+        for r in rules:
+            lines.append(f"- {r['check']}")
+    return "\n".join(lines)
+
+
+def on_package_change(code, stage):
+    return package_info_md(code), checklist_md(code, stage)
+
+
+def on_stage_change(code, stage):
+    return checklist_md(code, stage)
+
+
+def run_verification(code, stage, files):
+    """OCR the uploaded files and verify them against the checklist + STG rules.
+
+    Delegates to the Gemma 3 12B model (via OpenRouter) in llm.verify.
+    """
+    if not code:
+        return "⚠️ Please select a package code first."
+    if not stage:
+        return "⚠️ Please select a stage (Pre-Auth or Claim)."
+    if not files:
+        return "⚠️ Please upload at least one document to verify."
+
+    pkg = get_package(code)
+    docs = required_documents(code, stage)
+
+    # Gradio gives file paths (str) or objects with a .name attribute.
+    paths = [getattr(f, "name", f) for f in files]
+
+    try:
+        return verify(pkg, STAGE_LABELS[stage], docs, pkg.get("rules", []), paths)
+    except LLMConfigError as e:
+        return f"## ⚠️ Configuration error\n\n{e}"
+    except Exception as e:  # surface API/network errors to the user, don't crash the UI
+        return (
+            "## ❌ Verification failed\n\n"
+            f"The model call did not complete: `{type(e).__name__}: {e}`\n\n"
+            "Check your OpenRouter key/credits and network connection, then retry."
+        )
+
+
+with gr.Blocks(title="Hospital Claim Verification — PMJAY") as demo:
+    gr.Markdown(
+        "# 🏥 Hospital Claim Verification\n"
+        "Pre-check PMJAY **pre-authorization** and **claim submission** documents "
+        "against the Standard Treatment Guidelines — *before* uploading to the "
+        "Government portal. Catch compliance errors early and reduce rejections."
+    )
+
+    with gr.Row():
+        with gr.Column(scale=1):
+            package_dd = gr.Dropdown(
+                choices=dropdown_choices(),
+                label="1. Select Package Code",
+                info="PMJAY Health Benefit Package",
+            )
+            package_info = gr.Markdown()
+            stage_radio = gr.Radio(
+                choices=STAGE_RADIO_CHOICES,
+                label="2. Select Stage",
+                info="Pre-auth = before treatment · Claim = after treatment",
+            )
+            files_in = gr.File(
+                label="3. Upload documents",
+                file_count="multiple",
+                file_types=["image", ".pdf"],
+            )
+            verify_btn = gr.Button("Run Verification", variant="primary")
+
+        with gr.Column(scale=1):
+            checklist_out = gr.Markdown(
+                "_Select a package and stage to see the required documents._"
+            )
+
+    gr.Markdown("---")
+    result_out = gr.Markdown()
+
+    package_dd.change(
+        on_package_change,
+        inputs=[package_dd, stage_radio],
+        outputs=[package_info, checklist_out],
+    )
+    stage_radio.change(
+        on_stage_change,
+        inputs=[package_dd, stage_radio],
+        outputs=[checklist_out],
+    )
+    verify_btn.click(
+        run_verification,
+        inputs=[package_dd, stage_radio, files_in],
+        outputs=[result_out],
+    )
+
+
+if __name__ == "__main__":
+    demo.launch()

llm.py

@@ -0,0 +1,321 @@
+"""
+LLM verification — local, on-device (no cloud APIs).
+
+Runs Gemma 3 (a small, <=32B open model) through llama.cpp via llama-cpp-python:
+  1. OCR / read each uploaded file (images, or PDF pages rendered to images), and
+  2. Check the document set against the required-documents checklist and the STG
+     content rules for the selected package + stage.
+
+Model is configurable by environment variable so the same code runs a small
+model locally (e.g. Gemma 3 4B on a 16GB laptop) and the full model on the
+Space (Gemma 3 12B on 32GB):
+
+    CLAIMREADY_GGUF_REPO    default: ggml-org/gemma-3-12b-it-GGUF
+    CLAIMREADY_GGUF_FILE    default: gemma-3-12b-it-Q4_K_M.gguf
+    CLAIMREADY_MMPROJ_FILE  default: mmproj-model-f16.gguf
+    CLAIMREADY_N_CTX        default: 8192
+
+PDFs are rendered to images with PyMuPDF (no poppler), images are downscaled to
+keep vision encoding fast, and everything is fed to the model as images. The
+model is asked to return strict JSON, which we render into a readable verdict.
+"""
+
+import base64
+import io
+import json
+import os
+
+import fitz  # PyMuPDF
+from PIL import Image
+
+# Vision tiling: keep the longest side at/under this so each image is ~one tile.
+_MAX_IMAGE_SIDE = 896
+# Safety cap on total images per verification (PDF pages + image files).
+_MAX_IMAGES = 16
+# DPI used when rasterizing PDF pages.
+_PDF_DPI = 150
+
+_IMAGE_EXTS = {".png", ".jpg", ".jpeg", ".webp", ".gif", ".bmp", ".tiff", ".tif"}
+
+
+class LLMConfigError(RuntimeError):
+    """Raised when the model cannot be loaded."""
+
+
+# --- Model (loaded once, lazily) ----------------------------------------------
+
+_MODEL = None
+
+
+def _get_model():
+    global _MODEL
+    if _MODEL is not None:
+        return _MODEL
+    try:
+        from huggingface_hub import hf_hub_download
+        from llama_cpp import Llama
+        from llama_cpp.llama_chat_format import MTMDChatHandler
+
+        repo = os.environ.get("CLAIMREADY_GGUF_REPO", "ggml-org/gemma-3-12b-it-GGUF")
+        model_file = os.environ.get("CLAIMREADY_GGUF_FILE", "gemma-3-12b-it-Q4_K_M.gguf")
+        mmproj_file = os.environ.get("CLAIMREADY_MMPROJ_FILE", "mmproj-model-f16.gguf")
+        n_ctx = int(os.environ.get("CLAIMREADY_N_CTX", "8192"))
+
+        model_path = hf_hub_download(repo, model_file)
+        mmproj_path = hf_hub_download(repo, mmproj_file)
+        handler = MTMDChatHandler(clip_model_path=mmproj_path)
+        _MODEL = Llama(
+            model_path=model_path,
+            chat_handler=handler,
+            n_ctx=n_ctx,
+            n_threads=os.cpu_count(),
+            verbose=False,
+        )
+        return _MODEL
+    except Exception as e:  # surface a clear message to the UI
+        raise LLMConfigError(
+            f"Could not load the local model ({type(e).__name__}: {e}). "
+            "Check the CLAIMREADY_GGUF_* settings and that the model can be downloaded."
+        )
+
+
+# --- File -> image content parts ----------------------------------------------
+
+def _img_to_data_url(img):
+    """Downscale a PIL image to one vision tile and return a JPEG data URL."""
+    img = img.convert("RGB")
+    w, h = img.size
+    scale = min(1.0, _MAX_IMAGE_SIDE / max(w, h))
+    if scale < 1.0:
+        img = img.resize((max(1, int(w * scale)), max(1, int(h * scale))))
+    buf = io.BytesIO()
+    img.save(buf, format="JPEG", quality=90)
+    b64 = base64.b64encode(buf.getvalue()).decode("ascii")
+    return {"type": "image_url", "image_url": {"url": f"data:image/jpeg;base64,{b64}"}}
+
+
+def _file_to_parts(path):
+    """Turn an uploaded file into a list of image content parts."""
+    ext = os.path.splitext(path)[1].lower()
+    if ext == ".pdf":
+        parts = []
+        with fitz.open(path) as doc:
+            for page in doc:
+                pix = page.get_pixmap(dpi=_PDF_DPI)
+                img = Image.open(io.BytesIO(pix.tobytes("png")))
+                parts.append(_img_to_data_url(img))
+        return parts
+    if ext in _IMAGE_EXTS:
+        return [_img_to_data_url(Image.open(path))]
+    # Unknown extension: try to open as an image, else skip.
+    try:
+        return [_img_to_data_url(Image.open(path))]
+    except Exception:
+        return []
+
+
+# --- Prompt -------------------------------------------------------------------
+
+def _checklist_text(docs, rules):
+    lines = ["REQUIRED DOCUMENTS:"]
+    for d in docs:
+        lines.append(f"- id={d['id']} | {d['label']} — verify: {d['verify']}")
+    if rules:
+        lines.append("")
+        lines.append("CONTENT RULES (check against the text/values in the documents):")
+        for r in rules:
+            lines.append(f"- id={r['id']} | {r['check']}")
+    return "\n".join(lines)
+
+
+def _system_prompt():
+    return (
+        "You are a PMJAY (Ayushman Bharat) claims processing doctor reviewing a "
+        "hospital's pre-authorization / claim document set against the Standard "
+        "Treatment Guidelines BEFORE it is submitted to the government portal.\n\n"
+        "You will be given: (a) the package + stage, (b) a checklist of required "
+        "documents, (c) content rules to confirm, and (d) the uploaded files as "
+        "images (PDF pages are also given as images). Read every image carefully "
+        "(OCR the text).\n\n"
+        "For EACH required document, decide whether it is present among the "
+        "uploaded files and whether its content matches what must be verified. "
+        "For EACH content rule, decide pass / fail / unclear based on the text or "
+        "values you can read. Do not invent evidence — if you cannot read it, say "
+        "so and mark it unclear.\n\n"
+        "STRICT RULE EVALUATION — apply these without exception:\n"
+        "1. NUMERIC THRESHOLDS: When a rule states a numeric limit (e.g. '>= 101°F', "
+        "'>= 38.3°C', '< 7 g/dl', '> 55 years', '> 1 cm'), find the actual value in "
+        "the documents and compare it EXACTLY. Do NOT round up, approximate, or give "
+        "the benefit of the doubt. A reading of 100.8°F does NOT satisfy a '>= 101°F' "
+        "rule — that is a 'fail'. A hemoglobin of 7.2 does NOT satisfy '< 7 g/dl'.\n"
+        "2. DURATION / FREQUENCY: When a rule includes a time span (e.g. 'for more "
+        "than 2 days'), you must find explicit dated evidence covering that span "
+        "(e.g. fever documented on 3+ separate dated entries). A single reading never "
+        "satisfies a multi-day requirement — mark it 'unclear' if duration is not "
+        "documented.\n"
+        "3. PASS ONLY ON FULL MATCH: Mark 'pass' only when you can cite a specific, "
+        "readable value that meets EVERY part of the criterion (the value AND the "
+        "threshold AND any duration). If a value is present but below/outside the "
+        "bar, mark 'fail'. If you cannot read or locate the value, mark 'unclear'. "
+        "Never 'pass' by assumption.\n"
+        "4. UNITS: Convert units when needed to compare (°F<->°C, etc.), but always "
+        "state the actual value you read.\n"
+        "5. EVIDENCE: In every rule's 'evidence', quote the exact value you read and "
+        "state explicitly whether it meets the threshold (e.g. \"Temp 100.8°F read; "
+        "below the 101°F threshold -> fail\"). Keep document 'present' judgements "
+        "consistent with the evidence you cite for related rules.\n"
+        "6. STATUS MUST MATCH EVIDENCE: The 'status' you assign to a rule (and the "
+        "'present' boolean for a document) MUST agree with the conclusion written in "
+        "its own 'evidence'. If the evidence says the value is below/outside the bar, "
+        "or ends in '-> fail', the status is 'fail' — NEVER 'pass'. If the evidence "
+        "says the value cannot be read or located, the status is 'unclear'. Before "
+        "finalizing, re-read each evidence string and confirm the label agrees with "
+        "it. A 'pass' whose evidence concludes 'not met' is a contradiction and is "
+        "forbidden.\n\n"
+        "Respond with ONLY a JSON object (no markdown fences, no prose) matching:\n"
+        "{\n"
+        '  "documents": [{"id": str, "present": bool, "confidence": "high|medium|low", "evidence": str}],\n'
+        '  "rules": [{"id": str, "status": "pass|fail|unclear", "evidence": str}],\n'
+        '  "overall": "ready|not_ready|needs_review",\n'
+        '  "summary": str\n'
+        "}\n"
+        "Use the exact id values given. 'evidence' is a short phrase citing which "
+        "file and what you saw. Keep 'summary' to 1-3 sentences."
+    )
+
+
+def _build_user_content(pkg, stage_label, docs, rules, files):
+    """Build the user message: instructions + checklist text, then the images.
+
+    Gemma has no separate system role, so the system instructions are prepended
+    to the leading text block.
+    """
+    intro = (
+        f"{_system_prompt()}\n\n"
+        f"PACKAGE: {pkg['code']} — {pkg['name']} ({pkg.get('procedure', '')})\n"
+        f"STAGE: {stage_label}\n\n"
+        f"{_checklist_text(docs, rules)}\n\n"
+        f"UPLOADED FILES: the documents follow as images."
+    )
+    content = [{"type": "text", "text": intro}]
+
+    n_images = 0
+    truncated = False
+    for path in files:
+        if n_images >= _MAX_IMAGES:
+            truncated = True
+            break
+        for part in _file_to_parts(path):
+            if n_images >= _MAX_IMAGES:
+                truncated = True
+                break
+            content.append(part)
+            n_images += 1
+    return content, n_images, truncated
+
+
+def _parse_json(text):
+    """Best-effort parse of the model's JSON reply."""
+    text = (text or "").strip()
+    if text.startswith("```"):
+        text = text.strip("`")
+        if text.lower().startswith("json"):
+            text = text[4:]
+        text = text.strip()
+    try:
+        return json.loads(text)
+    except json.JSONDecodeError:
+        start, end = text.find("{"), text.rfind("}")
+        if start != -1 and end != -1 and end > start:
+            try:
+                return json.loads(text[start : end + 1])
+            except json.JSONDecodeError:
+                return None
+    return None
+
+
+def verify(pkg, stage_label, docs, rules, files):
+    """Run local LLM verification. Returns a markdown report string."""
+    model = _get_model()
+    content, _n_images, _truncated = _build_user_content(pkg, stage_label, docs, rules, files)
+
+    resp = model.create_chat_completion(
+        messages=[{"role": "user", "content": content}],
+        temperature=0,
+        max_tokens=1500,
+    )
+    raw = resp["choices"][0]["message"]["content"]
+    parsed = _parse_json(raw)
+    return _render(pkg, stage_label, docs, rules, parsed, raw)
+
+
+# --- Rendering -----------------------------------------------------------------
+
+_DOC_ICON = {True: "✅", False: "❌"}
+_RULE_ICON = {"pass": "✅", "fail": "❌", "unclear": "⚠️"}
+_OVERALL = {
+    "ready": "✅ **Ready to submit** — no blocking issues found.",
+    "not_ready": "❌ **Not ready** — fix the issues below before submitting.",
+    "needs_review": "⚠️ **Needs manual review** — some items could not be confirmed.",
+}
+
+
+def _render(pkg, stage_label, docs, rules, parsed, raw):
+    if parsed is None:
+        return (
+            "## ⚠️ Could not parse the model's response\n\n"
+            "The verification model replied, but not in the expected format. "
+            "Raw response below:\n\n"
+            f"```\n{(raw or '').strip()[:2000]}\n```"
+        )
+
+    rule_by_id = {r["id"]: r for r in rules}
+    res_docs = {d.get("id"): d for d in parsed.get("documents", []) if isinstance(d, dict)}
+    res_rules = {r.get("id"): r for r in parsed.get("rules", []) if isinstance(r, dict)}
+
+    lines = [
+        "## Verification result",
+        f"**{pkg['code']} — {pkg['name']}** · {stage_label}",
+        "",
+        _OVERALL.get(parsed.get("overall"), "⚠️ **Review the findings below.**"),
+    ]
+    if parsed.get("summary"):
+        lines += ["", f"> {parsed['summary']}"]
+
+    lines += ["", "### Documents"]
+    for d in docs:
+        r = res_docs.get(d["id"], {})
+        present = bool(r.get("present"))
+        icon = _DOC_ICON[present]
+        conf = r.get("confidence")
+        conf_txt = f" _(confidence: {conf})_" if conf else ""
+        lines.append(f"- {icon} **{d['label']}**{conf_txt}")
+        if r.get("evidence"):
+            lines.append(f"   - {r['evidence']}")
+        elif not r:
+            lines.append("   - _Not assessed by the model._")
+
+    if rules:
+        lines += ["", "### Content checks"]
+        for rule in rules:
+            r = res_rules.get(rule["id"], {})
+            status = r.get("status", "unclear")
+            icon = _RULE_ICON.get(status, "⚠️")
+            lines.append(f"- {icon} {rule['check']} — _{status}_")
+            if r.get("evidence"):
+                lines.append(f"   - {r['evidence']}")
+
+    missing = [d["label"] for d in docs if not res_docs.get(d["id"], {}).get("present")]
+    failed = [
+        rule_by_id[rid]["check"]
+        for rid, r in res_rules.items()
+        if r.get("status") == "fail" and rid in rule_by_id
+    ]
+    if missing or failed:
+        lines += ["", "### ⚠️ Action needed before submission"]
+        for m in missing:
+            lines.append(f"- Missing / unreadable document: **{m}**")
+        for f in failed:
+            lines.append(f"- Failed content check: {f}")
+
+    return "\n".join(lines)

packages.json

@@ -0,0 +1,446 @@
+{
+  "stages": {
+    "preauth": "Pre-Authorization (before treatment)",
+    "claim": "Claim Submission (after treatment)"
+  },
+  "packages": [
+    {
+      "code": "SG039A",
+      "name": "Cholecystectomy",
+      "procedure": "Without Exploration of CBD – Open",
+      "specialty": "General Surgery / Pediatric Surgery",
+      "price": "₹22,800",
+      "documents": {
+        "preauth": [
+          {
+            "id": "clinical_notes",
+            "label": "Clinical notes",
+            "verify": "Detailed history, signs & symptoms, and indication for the procedure"
+          },
+          {
+            "id": "usg_abdomen",
+            "label": "USG Upper Abdomen",
+            "verify": "Confirms presence of gall stone. If CBD exploration is planned, MRCP/USG should show large (>1cm) or multiple CBD stones"
+          },
+          {
+            "id": "lft",
+            "label": "LFT (Liver Function Test)",
+            "verify": "Report showing serum bilirubin, transaminase and alkaline phosphatase levels"
+          }
+        ],
+        "claim": [
+          {
+            "id": "operative_notes",
+            "label": "Operative notes",
+            "verify": "Detailed operative notes with indications and outcomes of the procedure"
+          },
+          {
+            "id": "pre_anesthesia",
+            "label": "Pre-anesthesia check-up report",
+            "verify": "Pre-anesthesia evaluation / fitness report"
+          },
+          {
+            "id": "discharge_summary",
+            "label": "Detailed Discharge Summary",
+            "verify": "Discharge summary including follow-up advice at the time of discharge"
+          },
+          {
+            "id": "intraop_photo",
+            "label": "Intra-operative photograph & gross specimen pictures",
+            "verify": "Intra-operative image and picture(s) of gross specimen removed. If CBD stones detected/suspected, photo/video of stones removed from CBD"
+          },
+          {
+            "id": "histopathology",
+            "label": "Histopathology report",
+            "verify": "Histopathology report of the specimen removed",
+            "note": "Can be submitted within 7 days of discharge"
+          }
+        ]
+      },
+      "rules": [
+        { "id": "usg_calculi", "check": "USG shows presence of calculi in the gall bladder" },
+        { "id": "pain_site", "check": "Patient has complaint of right hypochondrium or epigastric pain" },
+        { "id": "no_prior_chole", "check": "No prior cholecystectomy done in the past" }
+      ]
+    },
+    {
+      "code": "SG039B",
+      "name": "Cholecystectomy",
+      "procedure": "With Exploration of CBD – Open",
+      "specialty": "General Surgery / Pediatric Surgery",
+      "price": "₹22,800",
+      "documents": {
+        "preauth": [
+          {
+            "id": "clinical_notes",
+            "label": "Clinical notes",
+            "verify": "Detailed history, signs & symptoms, and indication for the procedure"
+          },
+          {
+            "id": "usg_abdomen",
+            "label": "USG Upper Abdomen",
+            "verify": "Confirms presence of gall stone. If CBD exploration is planned, MRCP/USG should show large (>1cm) or multiple CBD stones"
+          },
+          {
+            "id": "lft",
+            "label": "LFT (Liver Function Test)",
+            "verify": "Report showing serum bilirubin, transaminase and alkaline phosphatase levels"
+          }
+        ],
+        "claim": [
+          {
+            "id": "operative_notes",
+            "label": "Operative notes",
+            "verify": "Detailed operative notes with indications and outcomes of the procedure"
+          },
+          {
+            "id": "pre_anesthesia",
+            "label": "Pre-anesthesia check-up report",
+            "verify": "Pre-anesthesia evaluation / fitness report"
+          },
+          {
+            "id": "discharge_summary",
+            "label": "Detailed Discharge Summary",
+            "verify": "Discharge summary including follow-up advice at the time of discharge"
+          },
+          {
+            "id": "intraop_photo",
+            "label": "Intra-operative photograph & gross specimen pictures",
+            "verify": "Intra-operative image and picture(s) of gross specimen removed. If CBD stones detected/suspected, photo/video of stones removed from CBD"
+          },
+          {
+            "id": "histopathology",
+            "label": "Histopathology report",
+            "verify": "Histopathology report of the specimen removed",
+            "note": "Can be submitted within 7 days of discharge"
+          }
+        ]
+      },
+      "rules": [
+        { "id": "usg_calculi", "check": "USG shows presence of calculi in the gall bladder" },
+        { "id": "pain_site", "check": "Patient has complaint of right hypochondrium or epigastric pain" },
+        { "id": "no_prior_chole", "check": "No prior cholecystectomy done in the past" }
+      ]
+    },
+    {
+      "code": "SG039C",
+      "name": "Cholecystectomy",
+      "procedure": "Without Exploration of CBD – Lap.",
+      "specialty": "General Surgery / Pediatric Surgery",
+      "price": "₹22,800",
+      "documents": {
+        "preauth": [
+          {
+            "id": "clinical_notes",
+            "label": "Clinical notes",
+            "verify": "Detailed history, signs & symptoms, and indication for the procedure"
+          },
+          {
+            "id": "usg_abdomen",
+            "label": "USG Upper Abdomen",
+            "verify": "Confirms presence of gall stone. If CBD exploration is planned, MRCP/USG should show large (>1cm) or multiple CBD stones"
+          },
+          {
+            "id": "lft",
+            "label": "LFT (Liver Function Test)",
+            "verify": "Report showing serum bilirubin, transaminase and alkaline phosphatase levels"
+          }
+        ],
+        "claim": [
+          {
+            "id": "operative_notes",
+            "label": "Operative notes",
+            "verify": "Detailed operative notes with indications and outcomes of the procedure"
+          },
+          {
+            "id": "pre_anesthesia",
+            "label": "Pre-anesthesia check-up report",
+            "verify": "Pre-anesthesia evaluation / fitness report"
+          },
+          {
+            "id": "discharge_summary",
+            "label": "Detailed Discharge Summary",
+            "verify": "Discharge summary including follow-up advice at the time of discharge"
+          },
+          {
+            "id": "intraop_photo",
+            "label": "Intra-operative photograph & gross specimen pictures",
+            "verify": "Intra-operative image and picture(s) of gross specimen removed. If CBD stones detected/suspected, photo/video of stones removed from CBD"
+          },
+          {
+            "id": "histopathology",
+            "label": "Histopathology report",
+            "verify": "Histopathology report of the specimen removed",
+            "note": "Can be submitted within 7 days of discharge"
+          }
+        ]
+      },
+      "rules": [
+        { "id": "usg_calculi", "check": "USG shows presence of calculi in the gall bladder" },
+        { "id": "pain_site", "check": "Patient has complaint of right hypochondrium or epigastric pain" },
+        { "id": "no_prior_chole", "check": "No prior cholecystectomy done in the past" }
+      ]
+    },
+    {
+      "code": "SG039D",
+      "name": "Cholecystectomy",
+      "procedure": "With Exploration of CBD – Lap.",
+      "specialty": "General Surgery / Pediatric Surgery",
+      "price": "₹22,800",
+      "documents": {
+        "preauth": [
+          {
+            "id": "clinical_notes",
+            "label": "Clinical notes",
+            "verify": "Detailed history, signs & symptoms, and indication for the procedure"
+          },
+          {
+            "id": "usg_abdomen",
+            "label": "USG Upper Abdomen",
+            "verify": "Confirms presence of gall stone. If CBD exploration is planned, MRCP/USG should show large (>1cm) or multiple CBD stones"
+          },
+          {
+            "id": "lft",
+            "label": "LFT (Liver Function Test)",
+            "verify": "Report showing serum bilirubin, transaminase and alkaline phosphatase levels"
+          }
+        ],
+        "claim": [
+          {
+            "id": "operative_notes",
+            "label": "Operative notes",
+            "verify": "Detailed operative notes with indications and outcomes of the procedure"
+          },
+          {
+            "id": "pre_anesthesia",
+            "label": "Pre-anesthesia check-up report",
+            "verify": "Pre-anesthesia evaluation / fitness report"
+          },
+          {
+            "id": "discharge_summary",
+            "label": "Detailed Discharge Summary",
+            "verify": "Discharge summary including follow-up advice at the time of discharge"
+          },
+          {
+            "id": "intraop_photo",
+            "label": "Intra-operative photograph & gross specimen pictures",
+            "verify": "Intra-operative image and picture(s) of gross specimen removed. If CBD stones detected/suspected, photo/video of stones removed from CBD"
+          },
+          {
+            "id": "histopathology",
+            "label": "Histopathology report",
+            "verify": "Histopathology report of the specimen removed",
+            "note": "Can be submitted within 7 days of discharge"
+          }
+        ]
+      },
+      "rules": [
+        { "id": "usg_calculi", "check": "USG shows presence of calculi in the gall bladder" },
+        { "id": "pain_site", "check": "Patient has complaint of right hypochondrium or epigastric pain" },
+        { "id": "no_prior_chole", "check": "No prior cholecystectomy done in the past" }
+      ]
+    },
+    {
+      "code": "MG006A",
+      "name": "Enteric fever",
+      "procedure": "Enteric fever (medical management)",
+      "specialty": "General Medicine / Pediatric Medical Management",
+      "price": "₹1,800 (Gen) – ₹4,500 (ICU+vent)",
+      "documents": {
+        "preauth": [
+          {
+            "id": "clinical_notes",
+            "label": "Clinical notes with detailed history",
+            "verify": "Detailed clinical notes documenting history of fever and symptoms"
+          },
+          {
+            "id": "cbc_panel",
+            "label": "CBC, ESR, Peripheral smear, LFT",
+            "verify": "Baseline CBC, ESR, peripheral smear and LFT reports"
+          }
+        ],
+        "claim": [
+          {
+            "id": "indoor_case_papers",
+            "label": "Detailed Indoor case papers & treatment details",
+            "verify": "Indoor case papers with vitals and full treatment details"
+          },
+          {
+            "id": "post_cbc_panel",
+            "label": "Post-treatment CBC, ESR, Peripheral smear, LFT",
+            "verify": "Repeat CBC, ESR, peripheral smear and LFT after treatment"
+          },
+          {
+            "id": "discharge_summary",
+            "label": "Detailed Discharge Summary",
+            "verify": "Discharge summary including date of follow-up"
+          }
+        ]
+      },
+      "rules": [
+        { "id": "fever_threshold", "check": "Patient had fever >= 38.3°C / 101°F for more than 2 days (documented)" }
+      ]
+    },
+    {
+      "code": "MG064A",
+      "name": "Severe Anemia",
+      "procedure": "Severe anemia (medical management)",
+      "specialty": "General Medicine / Pediatric Medical Management",
+      "price": "₹1,800 (Gen) – ₹4,500 (ICU+vent)",
+      "documents": {
+        "preauth": [
+          {
+            "id": "clinical_notes",
+            "label": "Clinical notes",
+            "verify": "Evaluation findings, indications for the procedure, and planned line of treatment"
+          },
+          {
+            "id": "cbc_hb",
+            "label": "CBC, Hb report",
+            "verify": "Complete blood count and hemoglobin level report"
+          }
+        ],
+        "claim": [
+          {
+            "id": "indoor_case_papers",
+            "label": "Detailed Indoor case papers & treatment details",
+            "verify": "Indoor case papers with indications and treatment details. Should mention blood transfusion and ferrous sulphate injection"
+          },
+          {
+            "id": "post_cbc_hb",
+            "label": "Post-treatment CBC, Hb report",
+            "verify": "Repeat CBC / hemoglobin report after treatment"
+          },
+          {
+            "id": "discharge_summary",
+            "label": "Detailed Discharge Summary",
+            "verify": "Detailed discharge summary"
+          }
+        ]
+      },
+      "rules": [
+        { "id": "hb_below_7", "check": "Patient hemoglobin level was less than 7 g/dl" },
+        { "id": "transfusion", "check": "Patient underwent blood transfusion treatment" }
+      ]
+    },
+    {
+      "code": "SB039A",
+      "name": "Total Knee Replacement (Primary)",
+      "procedure": "Primary - Total Knee Replacement",
+      "specialty": "Orthopedics",
+      "price": "₹80,000",
+      "documents": {
+        "preauth": [
+          {
+            "id": "clinical_notes",
+            "label": "Clinical notes with indication for surgery",
+            "verify": "Clinical notes documenting the indication for knee replacement surgery"
+          },
+          {
+            "id": "knee_xray",
+            "label": "X-ray / CT of Knee (labelled)",
+            "verify": "X-ray or CT of the knee, labelled with patient ID, date and side (Left/Right)"
+          }
+        ],
+        "claim": [
+          {
+            "id": "indoor_case_papers",
+            "label": "Indoor case papers",
+            "verify": "Indoor case papers for the admission"
+          },
+          {
+            "id": "postop_photo",
+            "label": "Post-op clinical photograph",
+            "verify": "Post-operative clinical photograph"
+          },
+          {
+            "id": "postop_xray",
+            "label": "Post-op X-ray showing implant",
+            "verify": "Post-op X-ray of the operated knee showing the implant, labelled with patient ID, date and side (Left/Right)"
+          },
+          {
+            "id": "implant_invoice",
+            "label": "Implant invoice / barcode",
+            "verify": "Invoice or bar code of the implant used"
+          },
+          {
+            "id": "operative_note",
+            "label": "Detailed operative / procedure note",
+            "verify": "Detailed operative note"
+          },
+          {
+            "id": "discharge_summary",
+            "label": "Discharge Summary",
+            "verify": "Discharge summary"
+          }
+        ]
+      },
+      "rules": [
+        { "id": "postop_implant", "check": "Post-op X-ray of the knee shows presence of the implant" },
+        { "id": "primary_age", "check": "If indication is primary osteoarthritis (no trauma / systemic joint disease), patient age is > 55 years" }
+      ]
+    },
+    {
+      "code": "SB039B",
+      "name": "Total Knee Replacement (Revision)",
+      "procedure": "Revision - Total Knee Replacement",
+      "specialty": "Orthopedics",
+      "price": "₹1,30,000",
+      "documents": {
+        "preauth": [
+          {
+            "id": "clinical_notes",
+            "label": "Clinical notes with indication for surgery",
+            "verify": "Clinical notes documenting the indication for revision knee replacement"
+          },
+          {
+            "id": "knee_xray",
+            "label": "X-ray / CT of Knee (labelled)",
+            "verify": "X-ray or CT of the knee, labelled with patient ID, date and side (Left/Right)"
+          },
+          {
+            "id": "preop_xray_implant",
+            "label": "Pre-op X-ray showing existing implant",
+            "verify": "Pre-op X-ray of the affected knee showing the existing implant (required for revision cases)"
+          }
+        ],
+        "claim": [
+          {
+            "id": "indoor_case_papers",
+            "label": "Indoor case papers",
+            "verify": "Indoor case papers for the admission"
+          },
+          {
+            "id": "postop_photo",
+            "label": "Post-op clinical photograph",
+            "verify": "Post-operative clinical photograph"
+          },
+          {
+            "id": "postop_xray",
+            "label": "Post-op X-ray showing implant",
+            "verify": "Post-op X-ray of the operated knee showing the implant, labelled with patient ID, date and side (Left/Right)"
+          },
+          {
+            "id": "implant_invoice",
+            "label": "Implant invoice / barcode",
+            "verify": "Invoice or bar code of the implant used"
+          },
+          {
+            "id": "operative_note",
+            "label": "Detailed operative / procedure note",
+            "verify": "Detailed operative note"
+          },
+          {
+            "id": "discharge_summary",
+            "label": "Discharge Summary",
+            "verify": "Discharge summary"
+          }
+        ]
+      },
+      "rules": [
+        { "id": "postop_implant", "check": "Post-op X-ray of the knee shows presence of the implant" },
+        { "id": "preop_implant", "check": "Pre-op X-ray shows presence of the existing implant" }
+      ]
+    }
+  ]
+}

packages.py

@@ -0,0 +1,94 @@
+"""
+PMJAY package configuration — loaded from packages.json.
+
+packages.json is the single source of truth: each package carries the mandatory
+document checklist for both stages (pre-auth / claim) plus the "TMS rule" content
+checks the STG asks processing doctors to verify. Edit the JSON to add packages,
+documents, or rules — no code change required.
+
+The verdict engine treats `documents` as a deterministic presence checklist, and
+`rules` as content checks the LLM evaluates against OCR'd text.
+
+Source PDFs: stg_guidelines/STGs_ps1/
+"""
+
+import json
+import os
+
+_DATA_PATH = os.path.join(os.path.dirname(__file__), "packages.json")
+
+# Stage identifiers used across the app.
+STAGE_PREAUTH = "preauth"
+STAGE_CLAIM = "claim"
+
+
+def _load():
+    with open(_DATA_PATH, encoding="utf-8") as fh:
+        raw = json.load(fh)
+
+    stage_labels = raw.get("stages", {})
+
+    packages = {}
+    for pkg in raw.get("packages", []):
+        code = pkg["code"]
+        documents = {}
+        for stage, items in pkg.get("documents", {}).items():
+            documents[stage] = [
+                {
+                    "id": d["id"],
+                    "label": d["label"],
+                    "verify": d["verify"],
+                    "note": d.get("note", ""),
+                }
+                for d in items
+            ]
+        rules = [
+            {
+                "id": r["id"],
+                "check": r["check"],
+                "expected": r.get("expected", "Yes"),
+            }
+            for r in pkg.get("rules", [])
+        ]
+        packages[code] = {
+            "code": code,
+            "name": pkg["name"],
+            "procedure": pkg.get("procedure", ""),
+            "specialty": pkg.get("specialty", ""),
+            "price": pkg.get("price", ""),
+            "documents": documents,
+            "rules": rules,
+        }
+    return stage_labels, packages
+
+
+STAGE_LABELS, PACKAGES = _load()
+
+
+def dropdown_choices():
+    """Return [(label, code), ...] for the Gradio dropdown, grouped sensibly."""
+    # Count how many codes share each display name; only disambiguate with the
+    # procedure when a name is used by more than one package (e.g. the 4
+    # cholecystectomy variants).
+    name_counts = {}
+    for pkg in PACKAGES.values():
+        name_counts[pkg["name"]] = name_counts.get(pkg["name"], 0) + 1
+
+    choices = []
+    for code, pkg in PACKAGES.items():
+        label = f"{code} — {pkg['name']}"
+        if name_counts[pkg["name"]] > 1 and pkg.get("procedure"):
+            label += f" — {pkg['procedure']}"
+        choices.append((label, code))
+    return choices
+
+
+def get_package(code):
+    return PACKAGES.get(code)
+
+
+def required_documents(code, stage):
+    pkg = PACKAGES.get(code)
+    if not pkg:
+        return []
+    return pkg["documents"].get(stage, [])

requirements.txt

@@ -0,0 +1,6 @@
+--extra-index-url https://abetlen.github.io/llama-cpp-python/whl/cpu
+gradio>=4.44.0
+llama-cpp-python>=0.3.16
+pymupdf>=1.24.0
+pillow>=10.0.0
+huggingface_hub>=0.25.0