[Refactor] Decompose chat_service.py - 563 lines, 7 responsibilities

[GunaPalanivel]

Problem

chat_service.py is 563 lines and has accumulated 49 commits - nearly 2x the next highest-churn file in the repo. It owns at least 7 distinct responsibilities:

1. Query classification (_classify_query)
2. Query decomposition / sub-query splitting (_get_sub_queries)
3. Tool call routing (_get_agent_tool_calls)
4. Tool execution dispatch (_execute_search_tools)
5. Context retrieval and assembly (retrieve_context)
6. Relevance scoring and reformulation loop (_get_query_context_relevance)
7. Answer generation (generate_answer, get_chatbot_reply, get_chatbot_reply_stream)

It also re-defines CODE_BLOCK_PLACEHOLDER_PATTERN (line 36) which is already in api/tools/utils.py (line 17) - a leftover from when some logic was extracted but the constant was copy-pasted instead of imported.

Impact

• Every feature PR touches this file, causing merge conflicts across contributors
• Testing individual responsibilities requires mocking half the module
• The file is the single biggest contributor to the bus factor problem - one person has authored most of the 49 commits
• New contributors avoid touching it because a change anywhere can break everything

Proposed Decomposition

Split into focused modules under api/services/:

New Module	Responsibilities Moved
query_classifier.py	_classify_query, _get_sub_queries
tool_dispatcher.py	_get_agent_tool_calls, _execute_search_tools
context_retriever.py	retrieve_context, _get_query_context_relevance, reformulation loop
chat_service.py (slim)	get_chatbot_reply, get_chatbot_reply_stream, generate_answer - pure orchestration

Move CODE_BLOCK_PLACEHOLDER_PATTERN to a shared api/constants.py and import everywhere.

Migration Strategy

This can be done incrementally - one module at a time:

1. Extract query_classifier.py first (smallest, fewest dependencies)
2. Update imports in chat_service.py to use the extracted module
3. Run existing tests - they should pass without changes since the public API stays the same
4. Repeat for tool_dispatcher.py then context_retriever.py

Acceptance Criteria

• chat_service.py reduced to < 200 lines of orchestration logic
• Each extracted module has a clear single responsibility
• CODE_BLOCK_PLACEHOLDER_PATTERN defined in one place, imported everywhere
• All existing tests in tests/unit/services/test_chat_service.py pass without modification
• No changes to route-level imports (from api.services.chat_service import get_chatbot_reply)

References

• chatbot-core/api/services/chat_service.py (563 lines, 49 commits)
• chatbot-core/api/tools/utils.py line 17 (duplicate CODE_BLOCK_PLACEHOLDER_PATTERN)
• Related: [Enh] Reformulation loop in _get_reply_simple_query_pipeline() never mutates the query — same context retrieved on every iteration, fallback returned for all queries #191 (reformulation loop), [Bug] _execute_search_tools() crashes with TypeError — unguarded TOOL_REGISTRY miss and missing logger injection for 3/4 tools #190 (tool dispatch crashes), [Enh] get_chatbot_reply_new_architecture() is never called — agentic pipeline is unreachable dead code #189 (dead code)