feat(dsl): ANSWER verb - retrieval + pluggable reader LLM

[KailasMahavarkar]

Step 3 of retrieval-observability. Full-loop: retrieve with REMEMBER + synthesize via user-configured reader callable.

Shape

```python
def my_reader(prompt: str, max_tokens: int = 1000) -> str: ...

gs = GraphStore(reader=my_reader)

r = gs.execute('ANSWER "What is the capital of France?" LIMIT 3')

Result(

kind="answer",

data={

"answer": "Paris",

"cited_slots": ["n0", "n1", "n2"],

"candidates": [],

"reader": None,

},

count=1,

meta={"signals": },

)

```

Named registry for A/B'ing reader LLMs:

```python
gs = GraphStore(readers={"fast": a, "careful": b})
gs.execute('ANSWER "q" LIMIT 3 USING "careful"')
q.answer("q", limit=3, using="fast")
```

Grammar

```lark
answer_q: "ANSWER" STRING at_clause? tokens_clause? limit_clause? where_clause? using_reader?
using_reader: "USING" STRING
```

Mirrors `remember_q` shape + adds optional reader pick.

Zero LLM dep in core

graphstore ships no HTTP client, no litellm, no openai. Bring-your-own reader. Reader is a plain callable, not a config setting - held as a live reference on the executor.

Reader resolution order

1. `USING "name"` → `self._readers[name]` (raise if not found)
2. `self._reader` (default configured at construction)
3. Sole entry of `self._readers` if exactly one registered
4. None → `GraphStoreError`

Reader errors don't raise

Reader exceptions are caught. Result surfaces `data["error"]` with empty `answer`. Caller inspects without try/except on execute. Retrieval state is still returned so the caller can see what would've been fed to a working reader.

Tests

• `test_answer_end_to_end` - real retrieval + fake scripted reader
• `test_answer_without_reader_raises` - no reader configured
• `test_answer_picks_named_reader_via_using` - registry lookup
• `test_answer_unknown_named_reader_raises` - missing name
• `test_answer_reader_exception_surfaced_in_result` - isolation
• `test_answer_builder_roundtrip_matches_string_dsl` - builder parity
• `test_answer_builder_compiles_full_surface` - all clauses
• `test_answer_on_empty_store_still_calls_reader` - degenerate case

Full suite

1766 passed, 101 skipped, zero regressions.

What this enables

• LoCoMo + LongMemEval runners can replace their hand-rolled retrieve→format→LLM loops with a single `ANSWER` call.
• A/B over reader models becomes a one-line config.
• Attribution (which slots informed the answer) is first-class.
• Signal telemetry (Step 1) + plan dry-run (Step 2) + answer synthesis (Step 3) form a coherent observability + retrieval + synthesis triangle.

Next

Step 4: temporal anchor extraction. If the question contains a date ("in May 2024", "last Tuesday"), auto-add the `AT` clause. Targets temporal F1 - our weakest LoCoMo category.

🤖 Generated with Claude Code