feat(prompts): implement TOML-backed prompt loading and refactor message generation

python/ebook_search/answer.py

@@ -6,6 +6,7 @@ import logging
 from typing import TYPE_CHECKING
 
 from python.ebook_search.llm_interface import request_chat_completion
+from python.ebook_search.prompts import load_prompt
 
 if TYPE_CHECKING:
     import httpx
@@ -42,16 +43,7 @@ async def answer_query(
     content = await request_chat_completion(
         client,
         config,
-        [
+        load_prompt("answer").messages(query=query, context=context),
-            {
-                "role": "system",
-                "content": (
-                    "Answer only from the provided context. Cite sources with bracketed numbers like [1]. "
-                    "If the context is insufficient, say so."
-                ),
-            },
-            {"role": "user", "content": f"Question:\n{query}\n\nContext:\n{context}"},
-        ],
     )
 
     logger.info(f"ebook_answer_request_complete {config.chat_model=} answer_length={len(content)}")

python/ebook_search/prompts/__init__.py

@@ -0,0 +1,8 @@
+"""LLM prompt templates for EPUB search."""
+
+from python.ebook_search.prompts.lib import (
+    Prompt,
+    load_prompt,
+)
+
+__all__ = ["Prompt", "load_prompt"]

python/ebook_search/prompts/answer.toml

@@ -0,0 +1,9 @@
+system = """\
+Answer only from the provided context. Cite sources with bracketed numbers like [1]. \
+If the context is insufficient, say so."""
+user = """\
+Question:
+{query}
+
+Context:
+{context}"""

python/ebook_search/prompts/lib.py

@@ -0,0 +1,44 @@
+"""Load and render TOML-backed LLM prompt templates."""
+
+from __future__ import annotations
+
+import tomllib
+from dataclasses import dataclass
+from functools import cache
+from pathlib import Path
+
+
+@dataclass(frozen=True)
+class Prompt:
+    """A system and user prompt pair loaded from TOML."""
+
+    system: str
+    user: str
+
+    def messages(self, **values: str) -> list[dict[str, str]]:
+        """Render this prompt as OpenAI-style chat messages."""
+        return [
+            {"role": "system", "content": self.system.format(**values)},
+            {"role": "user", "content": self.user.format(**values)},
+        ]
+
+
+@cache
+def _get_prompt_dir() -> Path:
+    """Return the directory containing prompt template files."""
+    return Path(__file__).resolve().parent
+
+
+@cache
+def load_prompt(name: str) -> Prompt:
+    """Load and validate a named system and user prompt pair from TOML."""
+    path = _get_prompt_dir() / f"{name}.toml"
+    with path.open("rb") as file:
+        body = tomllib.load(file)
+
+    system = body.get("system")
+    user = body.get("user")
+    if not isinstance(system, str) or not isinstance(user, str):
+        msg = f"{path} must define string system and user prompts"
+        raise TypeError(msg)
+    return Prompt(system=system, user=user)

python/ebook_search/prompts/phrase_judge.toml

@@ -0,0 +1,10 @@
+system = """\
+Judge whether a candidate phrase from a book should be protected for RAG retrieval. \
+Do not extract new phrases. Reject common grammar fragments, ordinary nonspecific \
+phrases, unstable fragments, and phrases kept only because they are frequent. Keep \
+people, places, organizations, factions, events, technologies, fictional conditions, \
+magic systems, formal titles, named concepts, and recurring world-specific terms. \
+Return only a JSON object with keys: keep, canonical, category, aliases, confidence, \
+importance, allow_nested, suppress_children, reason."""
+
+user = "{candidate_json}"

python/ebook_search/protected_phrases/judge_ngrams.py

@@ -14,6 +14,7 @@ from sqlalchemy import select
 from sqlalchemy.ext.asyncio import AsyncSession
 
 from python.ebook_search.llm_interface import request_chat_completion
+from python.ebook_search.prompts import load_prompt
 from python.ebook_search.protected_phrases.extraction import (
     candidate_source_names,
     get_sample_contexts,
@@ -369,20 +370,7 @@ def build_judge_messages(candidate: PhraseCandidate) -> list[dict[str, str]]:
         "chapter_count": candidate.chapter_count,
         "contexts": candidate.sample_contexts,
     }
-    return [
+    return load_prompt("phrase_judge").messages(candidate_json=json.dumps(payload, ensure_ascii=True))
-        {
-            "role": "system",
-            "content": (
-                "Judge whether a candidate phrase from a book should be protected for RAG retrieval. "
-                "Do not extract new phrases. Reject common grammar fragments, ordinary nonspecific phrases, "
-                "unstable fragments, and phrases kept only because they are frequent. Keep people, places, "
-                "organizations, factions, events, technologies, fictional conditions, magic systems, formal titles, "
-                "named concepts, and recurring world-specific terms. Return only a JSON object with keys: keep, "
-                "canonical, category, aliases, confidence, importance, allow_nested, suppress_children, reason."
-            ),
-        },
-        {"role": "user", "content": json.dumps(payload, ensure_ascii=True)},
-    ]
 
 
 def parse_llm_judgment(content: str, config: EbookSearchConfig) -> LLMJudgment:

tests/ebook_search/test_http.py

@@ -93,6 +93,14 @@ async def test_answer_query_uses_httpx_chat_completions(mocker: MockerFixture) -
     payload = kwargs["json"]
     assert isinstance(payload, dict)
     assert payload["model"] == "deepseek-v4-flash"
+    assert payload["messages"] == [
+        {
+            "role": "system",
+            "content": "Answer only from the provided context. Cite sources with bracketed numbers like [1]. "
+            "If the context is insufficient, say so.",
+        },
+        {"role": "user", "content": "Question:\nquestion\n\nContext:\n[1] Book\nsource"},
+    ]
 
 
 async def test_embed_texts_uses_httpx_embeddings(mocker: MockerFixture) -> None: