Skip to content

Codebase intelligence for AI and humans: code health scores, auto-generated docs, git analytics, dead code detection, and architectural decisions via MCP.

fix(answer): stop the prompt formatter halving every page excerpt it fetches (#1168)

Swati Ahujacaf7e36612a0
diff --git a/packages/server/src/repowise/server/mcp_server/_answer_context.py b/packages/server/src/repowise/server/mcp_server/_answer_context.py
index 0fb1920c4..7af1323d3 100644
--- a/packages/server/src/repowise/server/mcp_server/_answer_context.py
+++ b/packages/server/src/repowise/server/mcp_server/_answer_context.py
@@ -94,7 +94,15 @@ _MAX_PRELUDE_COMMITS_PER_HIT = 2
# have to thread through the orchestrator.
_MAX_SYMBOL_DOC_CHARS = 120
_MATCHED_SYMBOL_DOC_CHARS = 400
+# A hit's fallback text is a one-line LLM summary or a 200-char page opener,
+# so 800 is already generous for those.
_MAX_CHARS_PER_HIT_SUMMARY = 800
+# A real page excerpt is a different thing and gets its own cap. It was
+# sharing the summary cap, which truncated fetched page content to less than
+# half of what the fetch had asked the database for. This value must stay at
+# or above the fetch size — tool_answer.answer checks that at import time,
+# since the fetch size lives in another module.
+_MAX_CHARS_PER_HIT_EXCERPT = 1500
def is_why_question(question: str) -> bool:
@@ -302,6 +310,7 @@ def build_context_block(
decisions: list[dict] | None = None,
*,
max_chars_per_hit: int = _MAX_CHARS_PER_HIT_SUMMARY,
+ max_chars_per_excerpt: int = _MAX_CHARS_PER_HIT_EXCERPT,
) -> str:
"""Format the LLM prompt body: prelude + decisions block + per-hit excerpts.
@@ -315,7 +324,7 @@ def build_context_block(
parts.append(prelude.rstrip())
if decisions:
parts.append(_format_decisions_block(decisions))
- parts.append(_format_hits_block(hits, max_chars_per_hit))
+ parts.append(_format_hits_block(hits, max_chars_per_hit, max_chars_per_excerpt))
return "\n\n".join(p for p in parts if p)
@@ -331,16 +340,24 @@ def _format_decisions_block(decisions: list[dict]) -> str:
return "\n".join(lines)
-def _format_hits_block(hits: list[dict], max_chars_per_hit: int) -> str:
+def _format_hits_block(
+ hits: list[dict],
+ max_chars_per_hit: int,
+ max_chars_per_excerpt: int = _MAX_CHARS_PER_HIT_EXCERPT,
+) -> str:
"""Format the per-hit excerpts. Mirrors the prior tool_answer behaviour."""
parts: list[str] = []
for i, h in enumerate(hits, start=1):
- # Prefer a real page excerpt when one was attached (the ambiguous-
- # retrieval / always-synthesize path enriches non-dominant hits with
- # actual page content so the LLM reads across the candidates, not just
- # one-line summaries). Falls back to the summary/snippet otherwise.
- body_src = h.get("excerpt") or h.get("summary") or h.get("snippet") or ""
- body = body_src[:max_chars_per_hit]
+ # Prefer a real page excerpt when one was attached, so the LLM reads
+ # the page's actual prose rather than a one-line summary of it. It is
+ # capped separately: the excerpt's size was already decided by the
+ # fetch, and re-applying the (smaller) summary cap here threw away
+ # content the query had paid for.
+ excerpt = h.get("excerpt") or ""
+ if excerpt:
+ body = excerpt[:max_chars_per_excerpt]
+ else:
+ body = (h.get("summary") or h.get("snippet") or "")[:max_chars_per_hit]
# Tag expanded hits so the LLM knows they didn't surface in retrieval
# directly — useful context when deciding how much weight to put on
# them in the answer.
diff --git a/packages/server/src/repowise/server/mcp_server/tool_answer/answer.py b/packages/server/src/repowise/server/mcp_server/tool_answer/answer.py
index 487f5c126..9d8dc0340 100644
--- a/packages/server/src/repowise/server/mcp_server/tool_answer/answer.py
+++ b/packages/server/src/repowise/server/mcp_server/tool_answer/answer.py
@@ -53,6 +53,9 @@ from sqlalchemy import delete, select
from repowise.core.persistence.database import get_session
from repowise.core.persistence.models import AnswerCache
from repowise.core.registry import mcp_tool_registry as mcp
+from repowise.server.mcp_server._answer_context import (
+ _MAX_CHARS_PER_HIT_EXCERPT,
+)
from repowise.server.mcp_server._answer_context import (
build_context_block as _build_context_block_v2,
)
@@ -112,6 +115,7 @@ from repowise.server.mcp_server.tool_answer.config import (
_ANSWER_SCHEMA_VERSION,
_DOMINANCE_RATIO,
_ENRICH_TOP_N_HITS,
+ _GATED_EXCERPT_CHARS,
_GATED_RETURN_HITS,
_HIGH_CONFIDENCE_SCORE_FLOOR,
_INLINE_BODY_MAX_LINES,
@@ -151,6 +155,18 @@ from repowise.server.mcp_server.tool_answer.synthesis import (
_log = logging.getLogger("repowise.mcp.answer")
+# The excerpt fetch and the prompt formatter each cap page content, and they
+# live in different modules — which is how the formatter came to discard more
+# than half of every excerpt the fetch had paid a database round-trip for.
+# Checked here, at import, because this is the only module that sees both.
+if _MAX_CHARS_PER_HIT_EXCERPT < _GATED_EXCERPT_CHARS:
+ raise RuntimeError(
+ "get_answer would truncate the page content it fetches: the prompt "
+ f"formatter caps an excerpt at {_MAX_CHARS_PER_HIT_EXCERPT} chars "
+ f"while the fetch asks for {_GATED_EXCERPT_CHARS}. Raise "
+ "_MAX_CHARS_PER_HIT_EXCERPT or lower _GATED_EXCERPT_CHARS."
+ )
+
# Always-synthesize flag. Default ON: synthesis runs for every retrieval and the
# post-synthesis grading cascade demotes confidence instead of the tool
# abstaining, so coverage matches a research assistant that answers every
diff --git a/tests/unit/server/mcp/test_answer_excerpt_cap.py b/tests/unit/server/mcp/test_answer_excerpt_cap.py
new file mode 100644
index 000000000..80337a505
--- /dev/null
+++ b/tests/unit/server/mcp/test_answer_excerpt_cap.py
@@ -0,0 +1,134 @@
+"""The formatter must not shrink a page excerpt below what the fetch bought.
+
+``_GATED_EXCERPT_CHARS`` decides how much page content a hit's excerpt is
+worth fetching from the database. The context-block formatter then applied
+its own, smaller cap to the same string, so half of every fetched excerpt
+was thrown away after being paid for. The two constants live in different
+modules, which is why the mismatch went unnoticed; the relationship is now
+checked at import time and asserted here.
+"""
+
+from __future__ import annotations
+
+from datetime import UTC, datetime
+from types import SimpleNamespace
+
+import pytest
+
+from repowise.core.persistence.models import Page
+
+_PAGE_IDS = ("file_page:pkg/alpha/one.py", "file_page:pkg/alpha/two.py")
+_NOW = datetime(2026, 3, 19, 12, 0, 0, tzinfo=UTC)
+
+
+async def _add_pages(factory, repo_id: str, content: str) -> None:
+ """Two file pages carrying `content` as their body."""
+ async with factory() as s:
+ for page_id in _PAGE_IDS:
+ s.add(
+ Page(
+ id=page_id,
+ repository_id=repo_id,
+ page_type="file_page",
+ title=page_id.removeprefix("file_page:"),
+ content=content,
+ target_path=page_id.removeprefix("file_page:"),
+ source_hash=page_id,
+ model_name="mock",
+ provider_name="mock",
+ generation_level=2,
+ confidence=0.9,
+ freshness_status="fresh",
+ created_at=_NOW,
+ updated_at=_NOW,
+ )
+ )
+ await s.commit()
+
+
+def _patch_pipeline(monkeypatch, answer_mod, page_id: str) -> None:
+ """An ambiguous hit pair (ratio 1.05) on a real Page row — the path that
+ already fetches page content today."""
+
+ async def _fake_retrieve(question, ctx):
+ return [
+ {"page_id": page_id, "score": 2.0, "_sources": {"fts"}},
+ {"page_id": "file_page:pkg/alpha/two.py", "score": 1.9, "_sources": {"fts"}},
+ ]
+
+ async def _fake_hydrate(hits, ctx, *, scope=None):
+ for h in hits:
+ h["target_path"] = h["page_id"].removeprefix("file_page:")
+ h["title"] = h["target_path"]
+ h["summary"] = "One-line summary of the page."
+ h["snippet"] = ""
+ h["page_type"] = "file_page"
+ return hits
+
+ monkeypatch.setattr(answer_mod, "_hybrid_retrieve", _fake_retrieve)
+ monkeypatch.setattr(answer_mod, "_hydrate_hits", _fake_hydrate)
+
+
+def _capture_prompt(monkeypatch, answer_mod, answer_text: str) -> list[str]:
+ """Record the user prompt handed to synthesis; short-circuit the LLM."""
+ seen: list[str] = []
+
+ class _Provider:
+ provider_name = "mock"
+ model_name = "mock-1"
+
+ async def generate(self, **kwargs):
+ return SimpleNamespace(content=answer_text)
+
+ async def _fake_synthesize(provider, system_prompt, user_prompt):
+ seen.append(user_prompt)
+ return answer_text, None
+
+ monkeypatch.setattr(answer_mod, "_resolve_provider_for_answer", lambda _p: _Provider())
+ monkeypatch.setattr(answer_mod, "synthesize", _fake_synthesize)
+ return seen
+
+
+async def test_whole_fetched_excerpt_survives_formatting(setup_mcp, factory, monkeypatch):
+ import repowise.server.mcp_server.tool_answer.answer as answer_mod
+ from repowise.server.mcp_server import get_answer
+ from repowise.server.mcp_server.tool_answer.config import _GATED_EXCERPT_CHARS
+
+ body = "abcdefghij" * 400 # 4000 chars — longer than any cap in play
+ await _add_pages(factory, setup_mcp, body)
+ _patch_pipeline(monkeypatch, answer_mod, _PAGE_IDS[0])
+ prompts = _capture_prompt(monkeypatch, answer_mod, "Alpha writes (pkg/alpha/one.py).")
+
+ await get_answer("how does the alpha module handle the write path")
+
+ assert prompts, "synthesis must run"
+ assert body[:_GATED_EXCERPT_CHARS] in prompts[0], (
+ f"the whole {_GATED_EXCERPT_CHARS}-char fetched excerpt must reach the prompt"
+ )
+
+
+async def test_a_hit_without_page_content_still_formats(setup_mcp, factory, monkeypatch):
+ """No page body is a normal state, not an error: the hit falls back to its
+ summary and the prompt is still built."""
+ import repowise.server.mcp_server.tool_answer.answer as answer_mod
+ from repowise.server.mcp_server import get_answer
+
+ await _add_pages(factory, setup_mcp, "")
+ _patch_pipeline(monkeypatch, answer_mod, _PAGE_IDS[0])
+ prompts = _capture_prompt(monkeypatch, answer_mod, "Alpha writes (pkg/alpha/one.py).")
+
+ await get_answer("how does the alpha module handle the write path")
+
+ assert prompts, "synthesis must run"
+ assert "One-line summary of the page." in prompts[0]
+
+
+def test_format_excerpt_cap_is_at_least_the_fetch_size():
+ """Guard the relationship directly, so an edit to either constant fails
+ here rather than silently halving the prompt."""
+ from repowise.server.mcp_server._answer_context import _MAX_CHARS_PER_HIT_EXCERPT
+ from repowise.server.mcp_server.tool_answer.config import _GATED_EXCERPT_CHARS
+
+ assert _MAX_CHARS_PER_HIT_EXCERPT >= _GATED_EXCERPT_CHARS