feat(spaces): Step 3 — VLM + Extract Space app.py, prompts shipped in core

Both ZeroGPU Spaces now run the contract core.clients expects:
predict(prompt, schema_json, image_path) -> clean JSON string at /predict.

- hitech-vlm: Qwen3.6-35B-A3B-FP8 via AutoModelForImageTextToText, eager
module-scope load, @spaces.GPU(duration=120), torch_dtype="auto",
enable_thinking=False, greedy decode, max_new_tokens=3072. Handles image+text
and text-only.
- hitech-extract: NuExtract3 (bf16), @spaces.GPU(duration=90), schema_json passed
through as NuExtract's `template`. Handles image+text and text-only.
- Each Space strips <think>/fences/prose to a clean JSON object so clients.py's
json.loads never burns the single retry on a stray character.
- requirements.txt: dropped `spaces` (platform pins it) and `gradio` (sdk base
image provides it); pinned transformers>=4.57. README frontmatter: dropped
sdk_version, added python_version "3.12".

Prompts relocated INTO the package (core/src/hitech_ai_core/prompts/<app>/v1.md)
so they ship in the wheel for git-subdirectory installs; authored real v1
prompts for all four apps; api.py gains load_prompt() (importlib.resources),
replacing the placeholder dict. Verified the .md files are bundled in the built
wheel. README §7 layout updated. 10 core tests still pass.

Also adds project-scoped Supabase MCP (.mcp.json) for the Step-4 hitech-ai-platform
project (auth done interactively by the user).

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

README.md

@@ -4,15 +4,20 @@ emoji: 📄
 colorFrom: green
 colorTo: blue
 sdk: gradio
-sdk_version: "5.0"
+python_version: "3.12"
 app_file: app.py
 pinned: false
 license: apache-2.0
 ---
 
-Structured extraction Space for Hi-Tech AI Platform.
+Structured extraction Space for Hi-Tech AI Platform — text→JSON and image→JSON
+(scanned POs, spec sheets, emails). Routed the high-frequency extraction here to
+keep the 36 B VLM's quota low.
 
-Model: `numind/NuExtract3`
-Hardware: ZeroGPU (set in Space settings)
+Model: `numind/NuExtract3` (4.5 B, ~9 GB)
+Hardware: ZeroGPU `large`; `@spaces.GPU(duration=90)`
 
-Populated in Step 3.
+API contract (called by `core.clients`):
+`predict(prompt: str, schema_json: str, image_path: str | None) -> str` (JSON),
+at `api_name="/predict"`. `schema_json` is passed through as NuExtract3's
+`template`.

app.py

@@ -1 +1,115 @@
-# Step 3: Extract Space — numind/NuExtract3 with @spaces.GPU
+"""hitech-extract — numind/NuExtract3 schema-constrained extraction Space.
+
+Serves the contract that `core/clients.py::_predict_default` expects:
+
+    predict(prompt: str, schema_json: str, image_path: str | None) -> str   # JSON
+
+NuExtract3 is template-driven: it takes a `template=` chat-template kwarg
+describing the fields to extract. We pass `schema_json` (the Pydantic
+`model_json_schema()`) straight through as that template. NuExtract's own DSL
+(`"verbatim-string"`, `"number"`, ...) differs from JSON Schema, but NuExtract is
+a Qwen3.5 fine-tune that follows JSON structure, and `core.clients` retry-once +
+Pydantic validation is the safety net. A JSON-Schema → DSL converter is deferred
+until the Step-6 eval set shows it is needed (YAGNI for v1).
+
+ZeroGPU notes (see huggingface-zerogpu skill): eager module-scope load; `import
+spaces` is unconditional and omitted from requirements.txt; greedy decode with
+thinking disabled for clean JSON.
+"""
+
+from __future__ import annotations
+
+import re
+
+import gradio as gr
+import spaces
+import torch
+from PIL import Image
+from transformers import AutoModelForImageTextToText, AutoProcessor
+
+MODEL_ID = "numind/NuExtract3"
+MAX_NEW_TOKENS = 2048
+
+# Eager module-scope load. NuExtract3 is small (~9 GB bf16) and ships its own
+# qwen3_5 modeling code via trust_remote_code.
+processor = AutoProcessor.from_pretrained(MODEL_ID, trust_remote_code=True)
+model = AutoModelForImageTextToText.from_pretrained(
+    MODEL_ID,
+    dtype=torch.bfloat16,
+    device_map="auto",
+    trust_remote_code=True,
+).eval()
+
+
+def _clean_json(text: str) -> str:
+    """Drop <think> blocks, code fences and prose; keep the JSON object.
+
+    `core/clients.py` does an unforgiving `json.loads` on the return value.
+    """
+    text = re.sub(r"<think>.*?</think>", "", text, flags=re.DOTALL)
+    fence = re.search(r"```(?:json)?\s*(.*?)```", text, flags=re.DOTALL)
+    if fence:
+        text = fence.group(1)
+    start, end = text.find("{"), text.rfind("}")
+    if start != -1 and end > start:
+        text = text[start : end + 1]
+    return text.strip()
+
+
+def _build_messages(prompt: str, image_path: str | None) -> list[dict]:
+    content: list[dict] = []
+    if image_path:
+        content.append({"type": "image", "image": Image.open(image_path).convert("RGB")})
+    content.append({"type": "text", "text": prompt})
+    return [{"role": "user", "content": content}]
+
+
+@spaces.GPU(duration=90)
+def predict(prompt: str, schema_json: str, image_path: str | None) -> str:
+    """Run one schema-constrained extraction and return a JSON string."""
+    messages = _build_messages(prompt, image_path)
+    inputs = processor.apply_chat_template(
+        messages,
+        add_generation_prompt=True,
+        tokenize=True,
+        return_dict=True,
+        return_tensors="pt",
+        template=schema_json,  # NuExtract3 schema-constrained extraction
+        enable_thinking=False,
+    ).to(model.device)
+
+    with torch.inference_mode():
+        generated = model.generate(
+            **inputs,
+            max_new_tokens=MAX_NEW_TOKENS,
+            do_sample=False,
+        )
+
+    generated = generated[:, inputs["input_ids"].shape[1] :]
+    text = processor.batch_decode(
+        generated,
+        skip_special_tokens=True,
+        clean_up_tokenization_spaces=False,
+    )[0]
+    return _clean_json(text)
+
+
+# gr.Interface exposes `fn` at api_name="/predict", which is what core.clients calls.
+demo = gr.Interface(
+    fn=predict,
+    inputs=[
+        gr.Textbox(label="prompt"),
+        gr.Textbox(label="schema_json"),
+        gr.Image(type="filepath", label="image_path"),
+    ],
+    outputs=gr.Textbox(label="json"),
+    title="Hi-Tech Extract",
+    description=(
+        "NuExtract3 — schema-constrained document→JSON extraction (text or scanned "
+        "image) for the Hi-Tech AI Platform. Returns a JSON string for core.clients "
+        "to Pydantic-validate."
+    ),
+)
+
+if __name__ == "__main__":
+    demo.launch()

requirements.txt

@@ -1,3 +1,9 @@
-# Pinned in Step 3
-gradio>=4.0
-spaces
+# hitech-extract — numind/NuExtract3 on ZeroGPU.
+# NOTE: do NOT add `spaces` here — the ZeroGPU platform pins its own copy and a
+# conflicting pin breaks the build (see huggingface-zerogpu skill).
+# torch / torchvision come from the ZeroGPU base image; pinning them here risks a
+# version fight. If the build errors on a missing vision dep, add it here.
+# gradio is provided by the `sdk: gradio` base image — do not pin it here.
+transformers>=4.57.0
+accelerate>=1.0
+pillow