fix(prisma): extract prompt qwen-silence restructure; notebook methods disclose 32B stage-3 escalation (refs P43)

notebooks/palliative_care_rfi.py

@@ -160,9 +160,12 @@ def _(mo):
     > project criteria. **Stage-3 full-text eligibility** runs locally
     > (qwen2.5:14b, 32k context, per-item markdown capped at 60k chars
     > with an explicit truncation marker — the criteria stay in
-    > window). Full-text retrieval: Unpaywall → PMC OA (S3) → Europe
-    > PMC → mirror fallback; items with no retrievable full text
-    > resolve `uncertain / unavailable-fulltext` rather than silently
+    > window); items the 14B persistently returned empty responses for
+    > were escalated to **qwen2.5:32b** (20k-char cap to fit its
+    > context), mirroring the stage-2 adjudication pattern. Full-text
+    > retrieval: Unpaywall → PMC OA (S3) → Europe PMC → mirror
+    > fallback; items with no retrievable full text resolve
+    > `uncertain / unavailable-fulltext` rather than silently
     > dropping. The flow diagram above reflects whatever stage the
     > pipeline has reached — it is computed live.
     """)

src/prisma/extract.py

@@ -28,13 +28,22 @@ if TYPE_CHECKING:
 log = logging.getLogger(__name__)
 
 
-_SYSTEM = """\
-You are extracting structured data from a clinical study for a
-systematic review. Fill the `extract_study_data` tool fields using
-values reported in the paper. If a field is not reported, return an
-empty string — never invent values. Preserve original units. Use the
-enum values exactly as specified; leave the field empty if no enum
-member fits.
+# One-sentence system message — longer system prompts make qwen models
+# served by Ollama return an EMPTY response (no text, no tool call);
+# same quirk screen.py and eligibility.py hit. Guidance lives in the
+# cached first user message.
+_SYSTEM = (
+    "You are extracting structured data from a clinical study for a systematic review."
+)
+
+_GUIDANCE = """\
+## Task
+
+Fill the `extract_study_data` tool fields using values reported in the
+paper (next message). If a field is not reported, return an empty
+string — never invent values. Preserve original units. Use the enum
+values exactly as specified; leave the field empty if no enum member
+fits.
 
 Put any noteworthy observations, caveats, or uncertainty about the
 extraction into `extraction_notes` (one to three sentences max).
@@ -48,7 +57,8 @@ def _build_call(project: Project, item_markdown: str, tool: LLMTool) -> LLMCall:
             LLMMessage(
                 role="user",
                 content=(
-                    "## Extraction template (reference)\n\n"
+                    _GUIDANCE
+                    + "\n## Extraction template (reference)\n\n"
                     + project.extraction_template
                 ),
                 cache=True,