arborist/tests/test_llm_context_base.py
russell@unturf.com 71507f9327
qa: pipe wikitext-base-v1 into LLM context (was verifier-only)
Both ask() (single-doc) and query() (multi-source) now run the
assembled context through aborist.wikitext.to_base() before message
construction. Hermes sees clean prose instead of raw wikitext, which:

- lets the model quote verbatim (raw [[Cloud Strife]] was unquotable
  in fluent prose; the model paraphrased to bare "Cloud Strife" and
  the verifier flagged that as UNGROUNDED)
- shrinks token bills ~43% on Wikipedia chunks (measured: 125,152 →
  71,771 chars on the FF7 main-character query)
- aligns LLM-input with what the verifier already strips, so the
  prose-vs-prose substring check is end-to-end consistent

Gated on policy["base_version"] (= wikitext-base-v1) so:
- governance_policy_hash flows the choice — prior answers cached
  under raw-wikitext policy retain distinct cache_keys, no migration
- operators can opt out (set base_version=None) for debugging
- environments without mwparserfromhell soft-fall to raw wikitext
  with no other behavior change

Tests (4): default policy carries base_version (tripwire), ask() user
messages have wikitext markers stripped, opt-out reverts to raw
wikitext, governance_hash differs with vs without base_version.
2026-04-28 17:08:49 -04:00

168 lines
5.4 KiB
Python

"""LLM context is wikitext-stripped before send.
Before this lands, runner.ask() and query.query() handed Hermes raw
[[wikitext]] markup. The model paraphrased it into clean prose then
labeled the result as a verbatim quote — verifier correctly flagged
those as UNGROUNDED. With base_version in policy, both paths run
to_base() on the assembled context before message construction; Hermes
sees prose so it CAN quote verbatim, and the verifier compares prose-
to-prose end-to-end (idempotent — verify_quotes also runs to_base).
These tests use StubClient to capture exactly what messages would be
sent to the LLM, asserting the wikitext markup is gone before the
model would see it.
"""
from __future__ import annotations
from pathlib import Path
import pytest
from aborist.document import Document
from aborist.ingest import ingest_source
from aborist.qa.client import StubClient
from aborist.qa.runner import DEFAULT_POLICY, ask
from aborist.store import connect
from aborist.wikitext import BASE_VERSION
WIKITEXT_DOC = (
"{{Refimprove|date=October 2009}}\n"
"[[File:foo.jpg|thumb|caption]]\n"
"''[[Final Fantasy VII]]'' follows [[protagonist]] [[Cloud Strife]], "
"a troubled mercenary.<ref>Smith 2010</ref> "
"He joins [[AVALANCHE]] to stop [[Shinra Electric Power Company|Shinra]]."
)
class _OneDocSource:
source_type = "llm_ctx_test"
def __init__(self, content: str):
self._doc = Document(
uri="https://example.com/ff7",
content=content,
source_type=self.source_type,
title="Test",
)
def iter_documents(self):
yield self._doc
def _ingest_one(db_path: Path, content: str) -> str:
conn = connect(db_path)
try:
ingest_source(conn, _OneDocSource(content))
row = conn.execute(
"SELECT document_root FROM documents WHERE document_uri=?",
("https://example.com/ff7",),
).fetchone()
finally:
conn.close()
return row["document_root"]
# ---------------------------------------------------------------------------
# DEFAULT_POLICY carries base_version
# ---------------------------------------------------------------------------
def test_default_policy_includes_base_version():
"""If base_version disappears from default policy, the LLM goes back
to seeing raw wikitext silently. Tripwire."""
assert DEFAULT_POLICY["base_version"] == BASE_VERSION
# ---------------------------------------------------------------------------
# Stripped context flows to the LLM (ask path)
# ---------------------------------------------------------------------------
def _user_messages_text(client: StubClient) -> str:
"""Concatenate every user-role message from the LAST captured call."""
last = client.calls[-1]
return "\n".join(m["content"] for m in last["messages"] if m["role"] == "user")
def test_ask_passes_base_form_to_llm(tmp_path):
db = tmp_path / "qa.db"
root = _ingest_one(db, WIKITEXT_DOC)
client = StubClient(answer="I don't know based on the provided document.")
conn = connect(db)
try:
ask(
conn,
document_root=root,
question="who is the protagonist?",
client=client,
model_id="stub-model",
)
finally:
conn.close()
sent = _user_messages_text(client)
# Wikitext markers MUST be gone.
assert "[[" not in sent
assert "]]" not in sent
assert "{{Refimprove" not in sent
assert "<ref>" not in sent
assert "File:" not in sent
# Topical content survives.
assert "Cloud Strife" in sent
assert "Final Fantasy VII" in sent
assert "AVALANCHE" in sent
def test_ask_with_base_version_disabled_passes_raw_wikitext(tmp_path):
"""Operator escape hatch: setting base_version=None reverts to raw
wikitext context. Confirms the gate is effective and reversible."""
db = tmp_path / "qa.db"
root = _ingest_one(db, WIKITEXT_DOC)
custom_policy = dict(DEFAULT_POLICY)
custom_policy["base_version"] = None
client = StubClient(answer="stub")
conn = connect(db)
try:
ask(
conn,
document_root=root,
question="who is the protagonist?",
client=client,
model_id="stub-model",
policy=custom_policy,
)
finally:
conn.close()
sent = _user_messages_text(client)
# Raw wikitext markers should now BE present in what the LLM saw.
assert "[[" in sent
assert "]]" in sent
# ---------------------------------------------------------------------------
# governance_policy_hash flows base_version (cache_key separation)
# ---------------------------------------------------------------------------
def test_base_version_changes_governance_hash(tmp_path):
"""Two policies that differ only in base_version produce different
governance_policy_hash values. This is what keeps prior cached
answers (under raw-wikitext policy) from satisfying lookups under
the new (stripped) policy on the same question/source — they remain
distinct cache_keys."""
from aborist.qa.keys import governance_policy_hash
p_with = dict(DEFAULT_POLICY)
p_with["base_version"] = "wikitext-base-v1"
p_without = dict(DEFAULT_POLICY)
p_without["base_version"] = None
h_with = governance_policy_hash(p_with)
h_without = governance_policy_hash(p_without)
assert h_with != h_without