Pablan, as it stands

Self-hosted knowledge management for SMEs: a split-screen Markdown editor
whose sections an LLM refines while you write, and RAG question answering
over the documents that result. FastAPI + Postgres/pgvector on the back,
SvelteKit on the front, everything OpenAI-compatible and self-hostable.

Squashed into a single commit; the development history stays local.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CA43ZJda8Rbp2hKXNy8f6b

backend/app/authoring/prompts.py

@@ -0,0 +1,81 @@
+"""Refinement prompt for the writing editor — natural language only.
+
+The model refines exactly ONE section of a document the user is writing. The
+rest of the document travels as prefix/suffix context so the section stays
+coherent with its surroundings, but the model regenerates ONLY the section —
+a large document is never re-emitted whole (FIM-style).
+
+The base texts (persona, rules, framings) are admin-editable: they come from
+`app/prompts/overrides.py::get_prompt`, which returns a DB override when one
+exists and the code default (`app/prompts/defaults.py`) otherwise.
+"""
+
+from app.llm.client import ChatMessage
+from app.prompts.overrides import get_prompt
+
+
+def render_refine_prompt(
+    section: str,
+    *,
+    prefix: str,
+    suffix: str,
+    persona: str | None,
+    hint: str | None,
+    context: str | None = None,
+    knowledge: list[str] | None = None,
+) -> list[ChatMessage]:
+    # A template may carry its own persona; otherwise the admin-editable default.
+    system_parts = [persona.strip() if persona else get_prompt("refine_persona")]
+    if hint:
+        system_parts.append(f"What this section should convey: {hint}")
+    if context:
+        # Background from the chat this capture came from, so the refinement
+        # is on-topic — but only as orientation, never a source of new facts.
+        system_parts.append(
+            f"Background (the conversation this document came from, for "
+            f"orientation only — do not invent facts from it): {context}"
+        )
+    system_parts.append(get_prompt("refine_rules"))
+
+    # The document being edited leads the user turn (prefix/suffix/section);
+    # the retrieved knowledge trails it, because it changes on every call and
+    # keeping it last leaves the stable prompt prefix reusable between calls.
+    user_parts: list[str] = []
+    if prefix.strip():
+        user_parts.append(
+            f"Text before the section (context only, do not repeat it):\n{prefix}"
+        )
+    if suffix.strip():
+        user_parts.append(
+            f"Text after the section (context only, do not repeat it):\n{suffix}"
+        )
+    user_parts.append(f"Refine only this section:\n{section}")
+    if knowledge:
+        # What the company has already documented elsewhere. It is grounding,
+        # not source material: it keeps terminology and facts consistent and
+        # lets the section point at related documents, but it must not be
+        # copied in or become a way to add facts the section's notes do not
+        # support.
+        joined = "\n\n".join(knowledge)
+        user_parts.append(f"{get_prompt('grounding_framing')}\n{joined}")
+
+    return [
+        {"role": "system", "content": "\n\n".join(system_parts)},
+        {"role": "user", "content": "\n\n".join(user_parts)},
+    ]
+
+
+def render_topic_summary_prompt(transcript: str) -> list[ChatMessage]:
+    """Condense a conversation into a short search topic (a few words)."""
+    return [
+        {"role": "system", "content": get_prompt("topic_summary")},
+        {"role": "user", "content": f"Conversation:\n{transcript}"},
+    ]
+
+
+def render_title_prompt(content_md: str) -> list[ChatMessage]:
+    """Suggest a concise document title from its written content."""
+    return [
+        {"role": "system", "content": get_prompt("title")},
+        {"role": "user", "content": f"Document:\n{content_md}"},
+    ]