PR-2 of fb-41 multi-hop RAG. Decompose + retrieve + synthesize 3-stage
pipeline가 `opts.multi_hop=true` 일 때 dispatch. Dynamic decide loop
는 PR-3.
- `AskOpts.multi_hop: bool` 필드 추가 + `impl Default for AskOpts`
도입 (HOTFIXES 2026-05-07 의 known limitation 해소). 9 explicit
init site 모두 `multi_hop: false` 추가 — Default 도입으로 향후
`..Default::default()` 점진 migrate 가능.
- `RagPipeline::ask` 의 entry 에 dispatcher 한 줄
(`if opts.multi_hop { return self.ask_multi_hop(...) }`).
- `RagPipeline::ask_multi_hop` 신규 method. 1) decompose LLM call
→ JSON array of strings parse, 2) 각 sub-query 로 retrieve +
chunk_id dedup pool, 3) score gate / no-chunks 가드, 4)
pack_context (single-pass 와 helper 공유), 5) synthesize LLM
call w/ MULTI_HOP_SYNTHESIZE_SYSTEM_PROMPT, 6) citation extract
+ Answer build. `prompt_template_version` = "rag-multi-hop-v1"
로 stamp — eval `compare` 가 single-pass vs multi-hop 분리.
- Prompt const 신규: MULTI_HOP_DECOMPOSE_SYSTEM_PROMPT +
MULTI_HOP_DECOMPOSE_USER_TEMPLATE + MULTI_HOP_SYNTHESIZE_SYSTEM_PROMPT
+ PROMPT_TEMPLATE_VERSION_MULTI_HOP + MULTI_HOP_MAX_SUB_QUERIES_DEFAULT.
- `kebab_core::RefusalReason::MultiHopDecomposeFailed` variant 신규.
Cascade: kebab-store-sqlite `refusal_reason_label` + kebab-tui `ask
refusal render` exhaustive match 갱신.
- `parse_decompose_response` + `strip_markdown_json_fence` helper —
markdown code fence (```json / ```) strip + JSON array of strings
parse + trim + drop empty + cap at MULTI_HOP_MAX_SUB_QUERIES_DEFAULT.
None 반환 시 caller 가 `MultiHopDecomposeFailed` refusal.
Tests (55 passing total, 8 신규):
- 6 unit (parse_decompose_response 의 bare array / fence variants /
garbage / cap / trim 회귀 핀).
- 2 integration: `ask_multi_hop_dispatches_and_decompose_garbage_refuses`
(decompose garbage → MultiHopDecomposeFailed + 정확히 1 LLM call) +
`ask_with_multi_hop_false_keeps_single_pass_path` (회귀 핀, 기존
caller 자동 backwards-compat).
Happy-path multi-hop (decompose 성공 → synthesize) 의 integration
test 는 ScriptedLm helper 가 PR-3 의 decide loop 와 함께 도입될
때 같이 추가. 현 `MockLanguageModel` 는 canned single response 라
2-LLM-call sequence 핀 불가.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Two issues from Task 7 review:
1. CRITICAL — call_tool "ask" arm called blocking ask handle from async
context. OllamaLanguageModel::new builds reqwest::blocking::Client
which creates+drops a tokio runtime → panic inside async. Fix:
tokio::task::spawn_blocking wrap. Also applied preemptively to
"search" arm (SqliteStore + Lance open are blocking IO too).
2. IMPORTANT — ask tool's retrieval mode hardcoded to Lexical (test
workaround for provider="none"); CLI default is Hybrid. Fix: add
`mode: Option<String>` field to AskInput, default Hybrid in handle,
test passes mode=Some("lexical") explicitly to keep test functional
on provider="none".
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Fourth (and final v1) tool — `ask` (input: query / optional session_id).
Multi-turn via optional session_id (kebab_app::ask_with_session_with_config),
single-shot via ask_with_config when None. Refusal (grounded:false) NOT
mapped to isError — agent branches on the wire payload's grounded flag.
AskOpts has no Default impl (must construct manually). Answer carries no
schema_version field (tagged inline via entry().or_insert_with, idempotent).
Mode defaulted to Lexical: reqwest::blocking::Client::build creates and
drops a tokio runtime, panicking inside async context — the empty-corpus
refusal test avoids this via spawn_blocking; the tool itself uses Lexical
as the default mode since MCP callers typically run without an embedding
provider configured.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
