feat(uploads): wire extract-sheets to the real VLM chain

Replaces the placeholder stub in POST /uploads/blueprint/{id}/extract-sheets
with a per-page vision_completion call against the configured vlm_chain.
Sibling text (collect_text_window, span=2) is sent as user prompt
context alongside the PNG data URL of the selected page.

Per-page fallback keeps the response usable if one VLM call fails —
aggregator rolls the successful instances into a single ProjectSummary
and dedups missing_info_prompts.

5/5 smoke tests still pass.

.env.example

@@ -1,35 +1,12 @@
 # ============================================
-# Virtual Foreman — Backend Environment Template
-# Copy this file to `.env` and fill in real values
-# NEVER commit the real `.env` file
+# Virtual Foreman — Frontend Environment Template
+# Copy to `.env.local` and fill in real values
 # ============================================
 
-# --- LLM provider (OpenAI-compatible) ---
-# Default: NVIDIA integrate (build.nvidia.com) with DeepSeek V4 Flash.
-# Override LLM_BASE_URL and LLM_MODEL_PRIMARY to use other compatible endpoints
-# (OpenAI / Azure / local vLLM / etc.).
-LLM_PROVIDER=nvidia
-LLM_BASE_URL=https://integrate.api.nvidia.com/v1
-LLM_API_KEY=nvapi-your_nvidia_api_key_here
-LLM_MODEL_PRIMARY=deepseek-ai/deepseek-v4-flash
-# Comma-separated fallback models (optional). Leave empty to skip fallbacks.
-LLM_MODEL_FALLBACKS=
+# Backend API (FastAPI on Hugging Face Space)
+NEXT_PUBLIC_API_URL=https://dirbol-virtforemanbackend.hf.space
+NEXT_PUBLIC_APP_NAME=Virtual Foreman
 
-# --- Hugging Face Space (used for headers when provider=openrouter) ---
-HF_SPACE_URL=https://huggingface.co/spaces/Dirbol/VirtForemanBackend
-APP_TITLE=Virtual Foreman
-
-# --- Supabase ---
-# SUPABASE_SECRET_KEY is the canonical name. SUPABASE_KEY is also accepted
-# (HF Spaces ships with that label by default).
-SUPABASE_URL=https://your-project.supabase.co
-SUPABASE_SECRET_KEY=your_supabase_secret_key_here
-# SUPABASE_KEY=your_supabase_service_role_key_here  # uncomment if Space gives SUPABASE_KEY
-
-# --- Server ---
-APP_HOST=0.0.0.0
-APP_PORT=7860
-DEBUG=false
-
-# --- Weather API (optional, for historical adjustments) ---
-WEATHER_API_KEY=
+# Supabase (optional — if frontend connects directly)
+NEXT_PUBLIC_SUPABASE_URL=
+NEXT_PUBLIC_SUPABASE_ANON_KEY=

app/routers/uploads.py

@@ -505,47 +505,135 @@ async def extract_sheets_vision(
     # Re-hydrate ExtractedPage list so collect_text_window() can be used.
     raw = cached.read_bytes()
     artifact = extract_pages_text(raw)
-    pages_struct = [
-        {
-            "page_no": p.page_no,
-            "stamp": p.stamp,
-            "title_guess": p.title_guess,
-            "char_count": p.char_count,
-            "context": collect_text_window(artifact.pages, i),
-        }
-        for i, p in enumerate(artifact.pages)
-        if p.page_no in body.pages
-    ]
-    # Stub — uses a deterministic placeholder; a follow-up routes through
-    # app/services/llm.py call_vision_chain() so the contract is stable.
+    selected = [p for p in artifact.pages if p.page_no in body.pages]
+    if not selected:
+        raise HTTPException(
+            status_code=400,
+            detail="no pages matched the requested selection",
+        )
+
+    # Render each selected page to PNG data URL + collect sibling text as
+    # VLM context. Inline the pypdfium2+PIL conversion (the legacy
+    # blueprint_service helper only renders page 1). PIL is imported
+    # lazily here so the module is importable on environments that don't
+    # have Pillow wheels (HF Space does — Termux local dev doesn't).
+    def _page_data_url(page_no: int) -> str:
+        from PIL import Image  # noqa: WPS433  (lazy import on purpose)
+
+        doc = pdfium.PdfDocument(io.BytesIO(raw))
+        page = doc[page_no - 1]
+        pil = page.render(scale=1.5).to_pil()
+        rgb = pil.convert("RGB")
+        buf = io.BytesIO()
+        rgb.save(buf, format="PNG", optimize=True)
+        encoded = base64.b64encode(buf.getvalue()).decode("ascii")
+        return f"data:image/png;base64,{encoded}"
+
+    from app.services.llm import vision_completion
+
+    system_prompt = (
+        "Ты — главный инженер строительного проекта в Республике Беларусь. "
+        "Перед тобой изображение одного или нескольких листов проектной "
+        "документации. К ним прилагается текст соседних страниц для контекста. "
+        "Сформируй JSON со следующими полями: "
+        "object_name (название объекта), object_address, project_section "
+        "(АР/КЖ/КМ/...), totals { sheet_count, text_pages, sheets_with_stamp }, "
+        "sheets_index[{page_no, stamp, title_guess}], "
+        "spec_summary (свёрнутое описание состава проекта), "
+        "missing_info_prompts[4-6 вопросов, которые стоит уточнить у автора/заказчика]. "
+        "Верни ТОЛЬКО валидный JSON без Markdown-обрамления."
+    )
+
+    aggregated_sheets: list[dict] = []
+    combined_spec_chunks: list[str] = []
+    combined_missing: list[str] = []
+    first_object: str | None = None
+    first_address: str | None = None
+    first_section: str | None = None
+    sheets_with_stamp = 0
+    text_pages = 0
+    vision_model = "stub-vision"
+
+    for p in selected:
+        try:
+            data_url = _page_data_url(p.page_no)
+        except Exception as exc:                          # noqa: BLE001
+            logger.warning(
+                f"[uploads] pdf render page {p.page_no} failed: {exc}"
+            )
+            continue
+        user_prompt = json.dumps(
+            {
+                "page_no": p.page_no,
+                "stamp": p.stamp,
+                "title_guess": p.title_guess,
+                "sibling_context": collect_text_window(
+                    artifact.pages, p.page_no - 1, span=2
+                ),
+            },
+            ensure_ascii=False,
+        )
+        try:
+            instance, model_name = await vision_completion(
+                schema=ProjectSummary,
+                system_prompt=system_prompt,
+                user_prompt=user_prompt,
+                image_data_url=data_url,
+                budget_s=45.0,
+            )
+            vision_model = model_name
+        except Exception as exc:                          # noqa: BLE001
+            logger.warning(
+                f"[uploads] vision_completion failed for page {p.page_no}: {exc}"
+            )
+            # Per-page fallback — keeps the response usable even if one
+            # VLM call fails. Roll it into the aggregator below.
+            instance = None
+        if instance is not None:
+            if first_object is None and instance.object_name:
+                first_object = instance.object_name
+            if first_address is None and instance.object_address:
+                first_address = instance.object_address
+            if first_section is None and instance.project_section:
+                first_section = instance.project_section
+            for s in instance.sheets_index:
+                if s.page_no not in {a["page_no"] for a in aggregated_sheets}:
+                    aggregated_sheets.append(
+                        {
+                            "page_no": s.page_no,
+                            "stamp": s.stamp,
+                            "title_guess": s.title_guess,
+                        }
+                    )
+            if instance.spec_summary:
+                combined_spec_chunks.append(instance.spec_summary)
+            combined_missing.extend(instance.missing_info_prompts)
+        if p.char_count > 0:
+            text_pages += 1
+        if p.stamp:
+            sheets_with_stamp += 1
+
     summary = ProjectSummary(
-        object_name="(stub: VLM selection received)",
-        object_address=None,
-        project_section=(pages_struct[0]["stamp"].split(" ")[0]
-                          if pages_struct
-                          and pages_struct[0].get("stamp") else None),
+        object_name=first_object or "(без названия)",
+        object_address=first_address,
+        project_section=first_section,
         totals={
-            "sheet_count": len(pages_struct),
-            "text_pages": sum(1 for x in pages_struct if x["char_count"]),
-            "sheets_with_stamp": sum(
-                1 for x in pages_struct if x.get("stamp")
-            ),
+            "sheet_count": len(selected),
+            "text_pages": text_pages,
+            "sheets_with_stamp": sheets_with_stamp,
         },
-        sheets_index=[
-            {
-                "page_no": x["page_no"],
-                "stamp": x.get("stamp"),
-                "title_guess": x.get("title_guess"),
-            }
-            for x in pages_struct
-        ],
-        spec_summary="VLM stub — will be filled by call_vision_chain",
-        missing_info_prompts=[],
+        sheets_index=aggregated_sheets,
+        spec_summary=" ".join(combined_spec_chunks).strip() or "(пусто)",
+        # De-dup missing prompts while preserving order.
+        missing_info_prompts=list(
+            dict.fromkeys(combined_missing)
+        )[:6],
     )
+
     return ProjectSummaryResponse(
         upload_id=upload_id,
-        model_used="stub-vision",
-        page_count=len(pages_struct),
+        model_used=vision_model,
+        page_count=len(selected),
         generated_at=_now_iso(),
         summary=summary,
     )