Planner and react: multi-clause routing, conversation context resolution, clarifying question fallback

common/llm_services/base_llm.py

@@ -1038,6 +1038,8 @@ def select_retriever_prompt(self):
     # Operator-customizable retrieval strategy for the react agent: the first
     # action, then each next action driven by what the previous result returned.
     _AGENTIC_AGENT_USER_DEFAULT = """\
+- If the question refers to something mentioned earlier (e.g. "that plan", "its report", "その計画", "この制度"), look it up in ## Conversation and use the full name in every retrieval call. If ## Conversation contains multiple candidates and it is unclear which one is meant, ask the user one short clarifying question instead of guessing — do not call any retrieval tool until clarified.
+- When calling an unstructured retriever (hybrid, contextual, similarity, community), pass only the sub-question for that specific part as a standalone search query in the user's language. Do not pass the full multi-part question, a part already covered by another step, or an unresolved reference from conversation history.
 - For most questions, make your FIRST action a vector search (graphrag__hybrid_search or graphrag__contextual_search) — it gives the broadest grounding. Skip it only when you are highly confident the question is a pure structured-data request (an exact count, an attribute/id lookup, a relationship traversal, or an aggregation over typed graph data) that a generated graph query fully answers on its own.
 - Let each observation drive the next action: if the passages you got back name specific entities or relationships you still need hard facts about, follow up with a structural query; if a result is thin, empty, or off-target, widen its parameters (top_k, num_hops) or switch method rather than repeating the same call.
 - Before answering, check that every part of the question is covered with the specific facts and figures it asks for; if a required value, table, or entity is still missing, retrieve again (widen top_k / num_hops or switch method) rather than answering vaguely or partially.
@@ -1074,6 +1076,7 @@ def agentic_agent_prompt(self):
 Plan mechanics (fixed):
 - A later step may depend on an earlier one: set depends_on and use arg_bindings to pull a value from a prior step's result, e.g. {"question": "S1.context.result"}.
 - Retrieval params (top_k, num_hops, community_level) are optional; omit them to use defaults, or set higher values when you expect a broad answer.
+- For each unstructured step, set args.question to that clause only as a standalone search query in the user's language. Do not pass the full multi-part question, a clause already assigned to another step, or an unresolved reference from conversation history.
 - The final step MUST have kind="answer" and tool="" (the orchestrator synthesizes the answer from gathered context); it should depend_on all retrieval steps.
 
 Decide which retrievals to include, how many, and in what order using the "Retrieval Strategy" below. Return ONLY the structured plan.
@@ -1088,8 +1091,10 @@ def agentic_agent_prompt(self):
     # Strategy (operator-customizable) — moved out of the fixed rules so it can
     # be tuned without touching the role / act model / plan mechanics.
     _AGENTIC_PLANNER_USER_DEFAULT = """\
+- If the question refers to something from earlier in the conversation, resolve it from ## Conversation before building the plan and substitute the full name in every step's args. If ## Conversation contains multiple candidates and it is unclear which one is meant, plan only a final answer step (no retrieval) that asks the user one short clarifying question.
+- When a question has multiple clauses, assign each clause to its own retrieval step. If another clause still needs passages or typed graph facts after one is covered, plan hybrid/community/structural for that clause too.
 - Prioritize including at least one vector search step (graphrag__hybrid_search or graphrag__contextual_search) unless you are highly confident the question is a pure structured-data request — an exact count, an attribute/id lookup, a relationship traversal, or an aggregation over typed graph data — that a generated graph query fully answers on its own. Whenever the answer could plausibly live in document text (what/why/how/describe/summarize, definitions, explanations, figures), include a vector search step. When unsure, include vector search.
-- Use BOTH kinds when a question needs facts from the graph AND supporting text; you may run several of each, in any order. When you use STRUCTURAL, pair it with a vector search step unless the question is a pure structured-data request.
+- Use BOTH structural and unstructured kinds when a question needs facts from the graph AND supporting text; you may run several of each, in any order. When you use STRUCTURAL, pair it with a vector search step unless the question is a pure structured-data request.
 - Prefer the smallest plan that will work. Trivial/greeting questions need only the final answer step.
 - Tabular / numeric questions (a specific value, a row, a column total, a ranking, or a year-over-year comparison from a table or chart): prefer graphrag__contextual_search or graphrag__hybrid_search with top_k>=10 (these return atomic table chunks that preserve full row/column structure); avoid graphrag__similarity_search alone; quote any specific table label, column header, year, or unit from the question (e.g. "ROE 2023"); for "compare X across years/regions/categories" set top_k>=15."""
 
@@ -1200,6 +1205,7 @@ def hyde_prompt(self):
 - **Quote exact values from the source.** Numbers, units, time periods, and named entities must appear verbatim — do not round, approximate, or translate units. Keep units in their original format, script, and language. For example, if the source says `1,234 km`, write `1,234 km`, not `767 miles` or `about 1,200 km`.
 - **For comparison or "which is the highest" questions, list each candidate's value before stating the conclusion.** Show the working — do not jump directly to a one-line answer.
 - **Score** each context for relevance and use only the high-scoring ones; do not invent additional logic.
+- **Multi-part questions:** answer each part from its matching context. Use structured-query results for typed graph facts; use document passages for "what does the report/document say." Do not quote a structured note as what a named report says. If retrieved passages are off-topic for a part, say that part is not in the retrieved documents.
 - **Cover** the relevant information, especially image references that carry critical visual information.
 - **Format** the answer in Markdown — titles, paragraphs, bulleted / numbered lists, images, and tables. Place images and tables below the related text section.
 - **Tables**: every row, including the header, starts on a new line.