Method decompose

rag_system/retrieval/query_transformer.py:10–292

Decompose *query* into standalone sub-queries. Parameters ---------- query : str The latest user message. chat_history : list[dict] | None Recent conversation turns (each item should contain at least the original user query under t

(self, query: str, chat_history: List[Dict[str, Any]] | None = None)

        self.llm_model = llm_model

    def decompose(self, query: str, chat_history: List[Dict[str, Any]] | None = None) -> List[str]:
        """Decompose *query* into standalone sub-queries.

        Parameters
        ----------
        query : str
            The latest user message.
        chat_history : list[dict] | None
            Recent conversation turns (each item should contain at least the original
            user query under the key ``"query"``). Only the **last 5** turns are
            included to keep the prompt short.
        """

        # ---- Limit history to last 5 user turns and extract the queries ----
        history_snippets: List[str] = []
        if chat_history:
            # Keep only the last 5 turns
            recent_turns = chat_history[-5:]
            # Extract user queries (fallback: full dict as string if key missing)
            for turn in recent_turns:
                history_snippets.append(str(turn.get("query", turn)))

        # Serialize chat_history for the prompt (single string)
        chat_history_text = " | ".join(history_snippets)

        # ---- Build the new SYSTEM prompt with added legacy examples ----
        system_prompt = """
You are an expert at query decomposition for a Retrieval-Augmented Generation (RAG) system.

Return one RFC-8259-compliant JSON object and nothing else.
Schema:
{
“requires_decomposition”: <bool>,
“reasoning”:              <string>,  // ≤ 50 words
“resolved_query”:         <string>,  // query after context resolution
“sub_queries”:            <string[]> // 1–10 standalone items
}

Think step-by-step internally, but reveal only the concise reasoning.

⸻

Context Resolution  (perform FIRST)

You will receive:
	•	query – the current user message
	•	chat_history – the most recent user turns (may be empty)

If query contains pronouns, ellipsis, or shorthand that can be unambiguously linked to something in chat_history, rewrite it to a fully self-contained question and place the result in resolved_query.
Otherwise, copy query into resolved_query unchanged.

⸻

When is decomposition REQUIRED?
	•	MULTI-PART questions joined by “and”, “or”, “also”, list commas, etc.
	•	COMPARATIVE / SUPERLATIVE questions (two or more entities, e.g. “bigger, better, fastest”).
	•	TEMPORAL / SEQUENTIAL questions (changes over time, event timelines).
	•	ENUMERATIONS (pros, cons, impacts).