mariano/nvi
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

recon-driven sql composer; pick → compose → execute; llm out of structural sql

Browse Source
This commit is contained in:
buenosairesam
2026-06-03 11:01:02 -03:00
parent 61494362a3
commit 29c620b2c2
27 changed files with 1516 additions and 249 deletions
Download Patch File Download Diff File Expand all files Collapse all files

49
api/analyses/direct_answer.py
View File

@@ -10,13 +10,15 @@ import logging
from typing import Any
from api import langfuse_client as lf
from api.analyses._pick import pick_for_question
from api.analyses.base import Analysis
from api.analyses.types import Finding
from api.composer import compose
from api.llm import chat
from api.prompts import load, render
from api.recon import load_recon
from api.runtime import events
from api.tools.execute_sql import execute_sql
from api.tools.text_to_sql import text_to_sql
logger = logging.getLogger("nvi.analyses.direct_answer")
@@ -29,23 +31,39 @@ class DirectAnswer(Analysis):
)
args_schema = {
"question": {"type": "string", "description": "Refined question this Analysis should answer."},
"hint_tables": {"type": "array", "items": "string", "description": "Tables the planner thinks are relevant (optional)."},
"allowed_metrics": {"type": "array", "items": "string", "description": "Metrics the planner judged relevant (optional; defaults to all)."},
}
async def run(self, args: dict[str, Any], question: str) -> Finding:
sub_q = args.get("question") or question
hint_tables = args.get("hint_tables")
allowed_metrics = args.get("allowed_metrics")
with lf.span("analysis.direct_answer", input={"question": sub_q}) as span:
try:
await events.publish_current(events.tool_call_start("text_to_sql", input={"question": sub_q}))
t2s = text_to_sql(sub_q, hint_tables=hint_tables)
recon = load_recon()
await events.publish_current(events.tool_call_start(
"pick_for_question", input={"question": sub_q},
))
outcome = pick_for_question(
sub_q, recon=recon, allowed_metrics=allowed_metrics,
span_name="direct_answer.pick",
)
await events.publish_current(events.tool_call_end(
"text_to_sql", output={"sql": t2s.sql, "tables": t2s.used_tables},
"pick_for_question", output={"pick": outcome.pick.to_dict()},
))
await events.publish_current(events.tool_call_start("execute_sql", input={"sql": t2s.sql}))
result = execute_sql(t2s.sql)
await events.publish_current(events.tool_call_start(
"compose_sql", input={"pick": outcome.pick.to_dict()},
))
composed = compose(outcome.pick, recon)
await events.publish_current(events.tool_call_end(
"compose_sql",
output={"sql": composed.sql, "tables": composed.used_tables},
))
await events.publish_current(events.tool_call_start("execute_sql", input={"sql": composed.sql}))
result = execute_sql(composed.sql)
await events.publish_current(events.tool_call_end(
"execute_sql",
output={"row_count": result.row_count, "truncated": result.truncated,
@@ -56,7 +74,7 @@ class DirectAnswer(Analysis):
interpret_user = render(
"direct_answer.interpret.user",
question=sub_q,
sql=t2s.sql,
sql=composed.sql,
rows=repr(result.as_dicts()[:20]),
)
await events.publish_current(events.llm_call(
@@ -64,7 +82,10 @@ class DirectAnswer(Analysis):
system_len=len(interpret_system),
user_len=len(interpret_user),
))
summary = chat(system=interpret_system, user=interpret_user, max_tokens=512)
summary = chat(
system=interpret_system, user=interpret_user, max_tokens=512,
span_name="direct_answer.interpret",
)
except Exception as e:
logger.exception("direct_answer failed")
await events.publish_current(events.tool_call_end("direct_answer", error=str(e)))
@@ -76,6 +97,10 @@ class DirectAnswer(Analysis):
analysis=self.name,
summary=summary,
rows=result.as_dicts()[:20],
sql=[t2s.sql],
metadata={"row_count": result.row_count, "truncated": result.truncated},
sql=[composed.sql],
metadata={
"row_count": result.row_count,
"truncated": result.truncated,
"metric": outcome.pick.metric,
},
)