Merge pull request #3 from athurlow/claude/qcal-copilot-mvp-OZ9wj

Fix CSV-as-plot rendering + surface VLM parse errors + force JSON output

src/qcal/analyzer.py

@@ -57,8 +57,21 @@ Only output the JSON. Do not wrap it in markdown fences.
 """
 
 USER_PROMPT_TEMPLATE = (
-    "Analyze this quantum calibration artifact ({source}) and return the JSON "
-    "described in the system prompt.{extra}"
+    "Analyze this quantum calibration artifact ({source}). "
+    "Respond with a SINGLE JSON object and nothing else — no prose before or "
+    "after, no markdown fences, no bullet points. The JSON must match this "
+    "schema exactly:\n\n"
+    "{{\n"
+    '  "experiment": "<string>",\n'
+    '  "qubit_id": "<string or null>",\n'
+    '  "issues": ["<string>", ...],\n'
+    '  "metrics": {{"<name>": "<value with units>", ...}},\n'
+    '  "recommended_parameters": {{"<name>": <number or string>, ...}},\n'
+    '  "drift_prediction": "<string>",\n'
+    '  "confidence": <float 0..1>,\n'
+    '  "notes": "<1-3 sentences>"\n'
+    "}}\n\n"
+    "Begin your reply with `{{` and end with `}}`.{extra}"
 )
 
 
@@ -79,6 +92,24 @@ class AnalysisResult:
     def markdown(self) -> str:
         if self.error:
             return f"**Analyzer error ({self.backend}):** {self.error}\n\n```\n{self.raw_text}\n```"
+        # The VLM returned 200 but the response wasn't parseable JSON (usually
+        # because the model wrote prose like "this doesn't look like a
+        # calibration plot"). Previously we'd render every field as 'n/a' with
+        # no indication of why — now surface the raw response so the user can
+        # see what the model actually said.
+        if not self.parsed:
+            snippet = (self.raw_text or "").strip() or "(empty response)"
+            if len(snippet) > 1200:
+                snippet = snippet[:1200] + "\n...[truncated]"
+            return (
+                f"**Analysis could not be parsed (backend: {self.backend}).** "
+                "The VLM returned a response but it wasn't valid JSON matching "
+                "the expected schema. Raw model output:\n\n"
+                f"```\n{snippet}\n```\n\n"
+                "_Tip: try a different image (Rabi/Ramsey/T1/T2/readout plots "
+                "work best), or re-run — the model occasionally flakes on "
+                "the first call after a cold start._"
+            )
         p = self.parsed
         lines = [
             f"**Experiment:** {p.get('experiment', 'n/a')}",
@@ -157,6 +188,10 @@ def _analyze_via_nim(image: Image.Image, extra: str, source: str) -> AnalysisRes
         ],
         "temperature": 0.2,
         "max_tokens": 1024,
+        # Force JSON-only output on backends that support OpenAI's response
+        # format parameter (vLLM, most NIM deployments). The VLM otherwise
+        # sometimes replies with markdown prose despite the system prompt.
+        "response_format": {"type": "json_object"},
     }
     try:
         resp = requests.post(

src/qcal/data.py

@@ -127,6 +127,62 @@ def _render_table_as_image(df: pd.DataFrame) -> Image.Image:
     return Image.open(buf).convert("RGB")
 
 
+def _render_csv_for_vlm(
+    df: pd.DataFrame,
+    *,
+    experiment_type: str = "unknown",
+    title: Optional[str] = None,
+) -> Image.Image:
+    """Render a user-uploaded CSV as whatever image the VLM can actually analyze.
+
+    The Ising Calibration VLM is trained on *plots* (Rabi traces, T1 decays,
+    IQ scatter, etc.), not on screenshots of numeric tables — feeding it a
+    table grid drops recognition confidence to ~0.2 and produces the "no clear
+    oscillations" failure mode. So for common CSV shapes we render a proper
+    line or scatter plot; only truly arbitrary tables fall back to the grid.
+    """
+    import matplotlib.pyplot as plt  # noqa: F401 — keeps mpl import local
+
+    numeric_cols = [c for c in df.columns if pd.api.types.is_numeric_dtype(df[c])]
+
+    # Two numeric columns: classic sweep (x, y). Covers Rabi / Ramsey / T1 / T2
+    # / resonator sweeps out of the box.
+    if len(numeric_cols) == 2:
+        x_col, y_col = numeric_cols
+        return _render_line(
+            df[y_col].to_numpy(),
+            df[x_col].to_numpy(),
+            experiment=experiment_type,
+            x_label=str(x_col),
+            y_label=str(y_col),
+            title=title,
+            fit=None,
+        )
+
+    # Readout IQ: two columns named like I/Q (any case, any order).
+    lower = {c.lower(): c for c in df.columns}
+    if "i" in lower and "q" in lower:
+        iq = df[[lower["i"], lower["q"]]].to_numpy()
+        return _render_scatter(iq, title=title)
+
+    # Single numeric column: plot vs row index.
+    if len(numeric_cols) == 1:
+        y_col = numeric_cols[0]
+        return _render_line(
+            df[y_col].to_numpy(),
+            None,
+            experiment=experiment_type,
+            x_label="sample index",
+            y_label=str(y_col),
+            title=title,
+            fit=None,
+        )
+
+    # Fall back to the table screenshot for wide/categorical tables the VLM
+    # probably can't interpret anyway.
+    return _render_table_as_image(df)
+
+
 def _fig_to_pil(fig) -> Image.Image:
     import matplotlib.pyplot as plt
 
@@ -449,7 +505,7 @@ def load_payload(
     if ext in SUPPORTED_TABLE_EXTS:
         sep = "," if ext == ".csv" else "\t"
         df = pd.read_csv(path, sep=sep)
-        img = _render_table_as_image(df)
+        img = _render_csv_for_vlm(df, experiment_type=experiment_type, title=name)
         return CalibrationPayload(image=img, table=df, source_name=name, kind="csv")
 
     if ext == ".npy":