From 76c3bec2ab1ab9ab01138cf84c55bd20db5163a4 Mon Sep 17 00:00:00 2001 From: Woody Date: Mon, 4 May 2026 17:22:14 +0800 Subject: [PATCH] feat: configurable SubQuestions via Step 1.2 system prompt page - Split 'Step 1: Query Decomposition' into Step 1.1 (prompt template) and Step 1.2 (format config with description + max_length) - Add create_subquestions_model() and parse_decompose_format() to decompose.py - QueryDecomposer reads decompose_format from DB, creates dynamic Pydantic model at runtime - PromptEditor renders Step 1.2 as textarea (description) + number input (max_length 1-5) - Graceful fallback to static SubQuestions when decompose_format unavailable --- backend/app/core/sqlite_db.py | 10 +- backend/app/models/decompose.py | 62 +++++++- backend/app/routers/prompts.py | 1 + backend/app/services/prompt_service.py | 1 + backend/app/services/query_decomposer.py | 36 ++++- frontend/src/components/PromptEditor.tsx | 144 +++++++++++++----- .../src/test/components/PromptEditor.test.tsx | 26 ++-- 7 files changed, 228 insertions(+), 52 deletions(-) diff --git a/backend/app/core/sqlite_db.py b/backend/app/core/sqlite_db.py index 73b1d96..a8b1e4e 100644 --- a/backend/app/core/sqlite_db.py +++ b/backend/app/core/sqlite_db.py @@ -11,12 +11,18 @@ logger = logging.getLogger(__name__) _SEED_DECOMPOSE = ( "Given this question: '{question}'\n\n" - "Break it down into 2-5 simplified sub-questions that would help " + "Break it down into 1-3 simplified sub-questions that would help " "search for relevant information. Each sub-question should be short " "and focused on one aspect.\n\n" 'Return a JSON array of strings: ["sub-question 1", "sub-question 2", ...]' ) +_SEED_DECOMPOSE_FORMAT = ( + '{"description": "請將問題/任務拆解成 1-3 個簡化子問題，標籤式主題必須清楚、簡潔、具體，' + '一看就明白（建議 3-8 個字），若涉及地點、地區、人物、時間、金額/財政 等關鍵資訊，必須包含在標籤中 。' + '具體提問/要求要精準、完整 並全部轉換成以下固定格式：\\n「標籤式主題：具體提問/要求」", "max_length": 3}' +) + _SEED_FILTER = ( "Given question '{question}' and these document chunks, rate each 0-10 for relevance. " "Return JSON array of scores.\n{chunks}\n" @@ -68,6 +74,7 @@ _SEED_PROFILES = [ _SEED_STEPS = [ "decompose", + "decompose_format", "filter", "generate", "generate_per_subq", @@ -77,6 +84,7 @@ _SEED_STEPS = [ ] _SEED_TEMPLATES = { "decompose": _SEED_DECOMPOSE, + "decompose_format": _SEED_DECOMPOSE_FORMAT, "filter": _SEED_FILTER, "generate": _SEED_GENERATE, "generate_per_subq": _SEED_GENERATE_PER_SUBQ, diff --git a/backend/app/models/decompose.py b/backend/app/models/decompose.py index d745e78..8c82647 100644 --- a/backend/app/models/decompose.py +++ b/backend/app/models/decompose.py @@ -1,11 +1,14 @@ -from pydantic import BaseModel, Field +from typing import Any + +from pydantic import BaseModel, Field, create_model class SubQuestions(BaseModel): - """Structured output model for query decomposition. + """Structured output model for query decomposition — static fallback. - Used with LangChain's with_structured_output() to guarantee - the LLM returns a valid list of sub-questions. + When ``decompose_format`` is available from PromptService, a dynamic + model with DB-configured description and max_length is used instead. + See ``create_subquestions_model()``. """ questions: list[str] = Field( @@ -13,3 +16,54 @@ class SubQuestions(BaseModel): min_length=1, max_length=3, ) + + +def create_subquestions_model(description: str, max_length: int) -> type[BaseModel]: + """Create a dynamic SubQuestions model with configurable field constraints. + + Args: + description: Field description text injected into the JSON format + instruction sent to the LLM. + max_length: Maximum number of sub-questions (1-5). + + Returns: + A Pydantic BaseModel subclass with the configured constraints. + """ + return create_model( + "SubQuestions", + questions=( + list[str], + Field( + description=description, + min_length=1, + max_length=max_length, + ), + ), + ) + + +def parse_decompose_format(raw: str) -> dict[str, Any]: + """Parse the decompose_format JSON string into a validated config dict. + + Returns a dict with ``description`` (str) and ``max_length`` (int). + Raises ``ValueError`` if the JSON is invalid or fields are missing/malformed. + """ + import json + + try: + config = json.loads(raw) + except json.JSONDecodeError as exc: + raise ValueError(f"decompose_format is not valid JSON: {exc}") from exc + + if not isinstance(config, dict): + raise ValueError("decompose_format must be a JSON object") + + description = config.get("description") + if not isinstance(description, str) or not description.strip(): + raise ValueError("decompose_format.description must be a non-empty string") + + max_length = config.get("max_length") + if not isinstance(max_length, int) or max_length < 1 or max_length > 5: + raise ValueError("decompose_format.max_length must be an integer between 1 and 5") + + return {"description": description.strip(), "max_length": max_length} diff --git a/backend/app/routers/prompts.py b/backend/app/routers/prompts.py index 6ef3555..f0bdf38 100644 --- a/backend/app/routers/prompts.py +++ b/backend/app/routers/prompts.py @@ -25,6 +25,7 @@ router = APIRouter(prefix="/api/v1/prompts", tags=["prompts"]) _VALID_NAMES = {"A", "B", "C"} _VALID_STEPS = { "decompose", + "decompose_format", "filter", "generate", "generate_per_subq", diff --git a/backend/app/services/prompt_service.py b/backend/app/services/prompt_service.py index 8025002..f5e8ecd 100644 --- a/backend/app/services/prompt_service.py +++ b/backend/app/services/prompt_service.py @@ -14,6 +14,7 @@ logger = logging.getLogger(__name__) _VALID_NAMES = {"A", "B", "C"} _VALID_STEPS = { "decompose", + "decompose_format", "filter", "generate", "generate_per_subq", diff --git a/backend/app/services/query_decomposer.py b/backend/app/services/query_decomposer.py index e7293e6..b537081 100644 --- a/backend/app/services/query_decomposer.py +++ b/backend/app/services/query_decomposer.py @@ -24,11 +24,21 @@ logger = logging.getLogger(__name__) # Fallback template used when prompt_service is not provided (tests, standalone). _BUILTIN_DECOMPOSE_TEMPLATE = ( "Given this question: '{question}'\n\n" - "Break it down into 2-5 simplified sub-questions that would help " + "Break it down into 1-3 simplified sub-questions that would help " "search for relevant information. Each sub-question should be short " "and focused on one aspect. Return as a JSON array of strings." ) +_DEFAULT_DECOMPOSE_FORMAT = { + "description": ( + "請將問題/任務拆解成 1-3 個簡化子問題，標籤式主題必須清楚、簡潔、具體，" + "一看就明白（建議 3-8 個字），若涉及地點、地區、人物、時間、金額/財政 等關鍵資訊，" + "必須包含在標籤中 。具體提問/要求要精準、完整 並全部轉換成以下固定格式：\n" + "「標籤式主題：具體提問/要求」" + ), + "max_length": 3, +} + def _extract_json_from_markdown(response: str) -> str: if not isinstance(response, str): @@ -98,12 +108,32 @@ class QueryDecomposer: prompt = template.replace("{question}", question) - from app.models.decompose import SubQuestions + from app.models.decompose import SubQuestions, create_subquestions_model, parse_decompose_format + + pydantic_model = SubQuestions + + if self._prompt_service is not None: + try: + format_raw = self._prompt_service.get_prompt_template("decompose_format") + format_config = parse_decompose_format(format_raw) + pydantic_model = create_subquestions_model( + description=format_config["description"], + max_length=format_config["max_length"], + ) + logger.info( + "Decompose format loaded from DB: max_length=%d", + format_config["max_length"], + ) + except Exception as exc: + logger.warning( + "Failed to load decompose_format from DB: %s. Falling back to static SubQuestions.", + exc, + ) try: result = await self.llm_client.complete_structured( prompt=prompt, - pydantic_model=SubQuestions, + pydantic_model=pydantic_model, step_name="QueryDecomposer", ) return result.questions, prompt diff --git a/frontend/src/components/PromptEditor.tsx b/frontend/src/components/PromptEditor.tsx index 58ce81a..0c8a592 100644 --- a/frontend/src/components/PromptEditor.tsx +++ b/frontend/src/components/PromptEditor.tsx @@ -14,7 +14,8 @@ export interface PromptEditorProps { } const STEPS = [ - { key: 'decompose', label: 'Step 1: Query Decomposition', placeholders: ['question'] }, + { key: 'decompose', label: 'Step 1.1: Query Decomposition', placeholders: ['question'] }, + { key: 'decompose_format', label: 'Step 1.2: Query Decomposition Format', placeholders: [], type: 'format_config' }, { key: 'filter_intro', label: 'Step 2.1: Filter Intro (Preamble)', placeholders: [] }, { key: 'filter_section', label: 'Step 2.2: Filter Section (Per Sub-Q)', placeholders: ['subq_idx', 'subq_question', 'chunks'] }, { key: 'filter_outro', label: 'Step 2.3: Filter Outro (Format)', placeholders: [] }, @@ -23,6 +24,7 @@ const STEPS = [ const VALID_PLACEHOLDERS: Record = { decompose: ['{question}'], + decompose_format: [], filter: ['{question}', '{chunks}'], generate: ['{question}', '{context}'], generate_per_subq: ['{context_sections}'], @@ -37,6 +39,68 @@ const findUnknownPlaceholders = (template: string, stepKey: string): string[] => return found.filter((p) => !valid.includes(p)) } +const DEFAULT_FORMAT_CONFIG = { description: '', max_length: 3 } + +function parseFormatConfig(raw: string): { description: string; max_length: number } { + try { + const parsed = JSON.parse(raw) + return { + description: typeof parsed.description === 'string' ? parsed.description : '', + max_length: typeof parsed.max_length === 'number' ? parsed.max_length : 3, + } + } catch { + return { ...DEFAULT_FORMAT_CONFIG } + } +} + +function serializeFormatConfig(description: string, max_length: number): string { + return JSON.stringify({ description, max_length }) +} + +interface FormatConfigFieldsProps { + value: string + disabled: boolean + onChange: (description: string, max_length: number) => void +} + +const FormatConfigFields: React.FC = ({ value, disabled, onChange }) => { + const config = parseFormatConfig(value) + + return ( +
+
+ +