bstein/atlasbot
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

atlasbot: skip deep retrieval for quick

Browse Source
This commit is contained in:
Brad Stein 2026-02-04 11:00:38 -03:00
parent 83e6a0fac2
commit f5e928fa59
1 changed files with 51 additions and 45 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

96
atlasbot/engine/answerer.py
View File

@ -78,6 +78,7 @@ class ModePlan:
chunk_group: int chunk_group: int
parallelism: int parallelism: int
score_retries: int score_retries: int
use_deep_retrieval: bool
use_tool: bool use_tool: bool
use_critic: bool use_critic: bool
use_gap: bool use_gap: bool
@ -360,57 +361,59 @@ class AnswerEngine:
) )
metric_facts: list[str] = [] metric_facts: list[str] = []
if classify.get("question_type") in {"metric", "diagnostic"} or force_metric: if classify.get("question_type") in {"metric", "diagnostic"} or force_metric:
if observer:
observer("retrieve", "extracting fact types")
fact_types = await _extract_fact_types(
call_llm,
normalized,
keyword_tokens,
plan,
)
if observer:
observer("retrieve", "deriving signals")
signals = await _derive_signals(
call_llm,
normalized,
fact_types,
plan,
)
if isinstance(signals, list):
signal_tokens = [str(item) for item in signals if item]
all_tokens = _merge_tokens(signal_tokens, keyword_tokens, question_tokens) all_tokens = _merge_tokens(signal_tokens, keyword_tokens, question_tokens)
if observer: if plan.use_deep_retrieval:
observer("retrieve", "scanning chunks")
candidate_lines: list[str] = []
if signals:
for chunk in selected:
chunk_lines = chunk["text"].splitlines()
if not chunk_lines:
continue
hits = await _scan_chunk_for_signals(
call_llm,
normalized,
signals,
chunk_lines,
plan,
)
if hits:
candidate_lines.extend(hits)
candidate_lines = list(dict.fromkeys(candidate_lines))
if candidate_lines:
if observer: if observer:
observer("retrieve", "pruning candidates") observer("retrieve", "extracting fact types")
metric_facts = await _prune_metric_candidates( fact_types = await _extract_fact_types(
call_llm, call_llm,
normalized, normalized,
candidate_lines, keyword_tokens,
plan, plan,
plan.metric_retries,
) )
if metric_facts: if observer:
key_facts = _merge_fact_lines(metric_facts, key_facts) observer("retrieve", "deriving signals")
if self._settings.debug_pipeline: signals = await _derive_signals(
_debug_log("metric_facts_selected", {"facts": metric_facts}) call_llm,
normalized,
fact_types,
plan,
)
if isinstance(signals, list):
signal_tokens = [str(item) for item in signals if item]
all_tokens = _merge_tokens(signal_tokens, keyword_tokens, question_tokens)
if observer:
observer("retrieve", "scanning chunks")
candidate_lines: list[str] = []
if signals:
for chunk in selected:
chunk_lines = chunk["text"].splitlines()
if not chunk_lines:
continue
hits = await _scan_chunk_for_signals(
call_llm,
normalized,
signals,
chunk_lines,
plan,
)
if hits:
candidate_lines.extend(hits)
candidate_lines = list(dict.fromkeys(candidate_lines))
if candidate_lines:
if observer:
observer("retrieve", "pruning candidates")
metric_facts = await _prune_metric_candidates(
call_llm,
normalized,
candidate_lines,
plan,
plan.metric_retries,
)
if metric_facts:
key_facts = _merge_fact_lines(metric_facts, key_facts)
if self._settings.debug_pipeline:
_debug_log("metric_facts_selected", {"facts": metric_facts})
if not metric_facts: if not metric_facts:
if observer: if observer:
observer("retrieve", "fallback metric selection") observer("retrieve", "fallback metric selection")
@ -1074,6 +1077,7 @@ def _mode_plan(settings: Settings, mode: str) -> ModePlan:
chunk_group=4, chunk_group=4,
parallelism=4, parallelism=4,
score_retries=3, score_retries=3,
use_deep_retrieval=True,
use_tool=True, use_tool=True,
use_critic=True, use_critic=True,
use_gap=True, use_gap=True,
@ -1092,6 +1096,7 @@ def _mode_plan(settings: Settings, mode: str) -> ModePlan:
chunk_group=4, chunk_group=4,
parallelism=2, parallelism=2,
score_retries=2, score_retries=2,
use_deep_retrieval=True,
use_tool=True, use_tool=True,
use_critic=True, use_critic=True,
use_gap=True, use_gap=True,
@ -1109,6 +1114,7 @@ def _mode_plan(settings: Settings, mode: str) -> ModePlan:
chunk_group=5, chunk_group=5,
parallelism=1, parallelism=1,
score_retries=1, score_retries=1,
use_deep_retrieval=False,
use_tool=False, use_tool=False,
use_critic=False, use_critic=False,
use_gap=False, use_gap=False,