Replaces the rigid FSM after Intro/AskLanguage with a single LLM call per
turn that sees:
* persona (what's already known — slots not re-asked)
* recent 10 chat_log messages (history)
* top-5 kei-sage atoms relevant to user_text (graph-RAG, not embeddings)
* raw user_text
LLM returns JSON {slot_updates, response_text, done, focus} which drives
the next state + persona patch + reply. No embeddings, no vector store —
kei-sage's FTS5 + Obsidian-style atom graph is the retrieval layer.
New files:
* src/retrieval.rs (101 LOC) — retrieve_context(chat_log, topics,
chat_id, query, history_n, atoms_k) -> RetrievalContext
* src/conversational.rs (157 LOC) — conversational_step
(state, persona, context, text, extractor, lang) -> StepOutput
Modified:
* src/serve.rs::run_fsm — branch on state: Intro/AskLanguage still go
through legacy handle_step (jump-start); everything else routes to
conversational_step with retrieval context.
* src/lib.rs — module declarations.
Tests (5 new, 60 total passing):
* parses_well_formed_llm_response
* done_true_transitions_to_ready
* invalid_json_falls_back_gracefully
* retrieve_returns_empty_on_empty_stores
* retrieve_finds_seeded_data
Verify:
* cargo check -p kei-buddy: PASS
* cargo test -p kei-buddy --lib: 60/0 (was 55, +5)
Why graph-RAG instead of embeddings: kei-sage already in tree (atoms +
edges + BFS + PageRank + FTS5). Explicit edges (message → topic →
contact) beat opaque cosine similarity for personal-assistant memory
where relationships are typed. No sqlite-vec dep, no embedding cost.
NOT deployed yet — needs server rebuild.
