Add a body attribute tier so query seeding can reach non-extraction content (#3313)

CHANGELOG.md

@@ -4,6 +4,8 @@ Full release notes with details on each version: [GitHub Releases](https://githu
 
 ## 0.9.63 (2026-09-16)
 
+- Feature: query seeding now considers a node's `body` attribute, not just its label, source path, and the curated `rationale` attribute. A non-extraction emitter (a Confluence page, a Jira issue) has nowhere else to put its content — storing it on an arbitrary attribute bought nothing on its own, since the node could never become a seed and was therefore never traversed no matter how relevant. `body` is the documented convention, matched at its own tier well below `rationale` since raw scraped prose is noisier than curated why-text (#3313, thanks @vongohren).
+
 - Feature: Elixir `alias`/`import`/`require`/`use` targets now resolve onto the module's `defmodule` node across files, so the internal module dependency graph is no longer dropped as dangling. Only top-level modules are indexed (a nested `defmodule`, labeled with its bare inner name, cannot capture an unrelated `use <Name>` from another file), and a same-file reference is left unresolved so it cannot clobber the structural `contains` edge (#3603, thanks @ayushcodes10).
 - Feature: a Rust `self.method()` call now resolves to a method defined on the same type in another file (the common split-`impl`-block layout), pooling methods across every `impl` of one type and refusing to link when two unrelated types share a bare name (#3602, thanks @ayushcodes10).
 - Feature: a Ruby member call `obj.foo` on a known-type receiver now resolves to a method `foo` inherited from a superclass, including across files, using the same conservative promotion as the implicit-self resolver — a single owning class, matching method kind, and one unambiguous ancestry chain, or it stays dangling (#3585, thanks @oleksii-tumanov).

graphify/serve.py

@@ -312,6 +312,17 @@ def _query_terms(question: str) -> list[str]:
 # toward term coverage, so a long rationale adds recall without winning back
 # an exact-label tier it did not earn.
 _RATIONALE_MATCH_BONUS = 0.75
+# A node emitted by something other than graphify's own extraction (a
+# Confluence page, a Jira issue) has nowhere to put its free text except an
+# arbitrary attribute, and nothing considers those for seeding: the content
+# is in the graph, the ranker is fine, the node is just never visited (#3313).
+# `body` is the documented convention such an emitter can use. Weighted well
+# below the curated `rationale` tier (which the extraction spec itself
+# writes) rather than reusing it, since raw scraped/pasted text is noisier
+# and an unweighted body match was measured to make retrieval WORSE - a few
+# hundred characters of prose matches many query terms, so a page that
+# merely mentions the words ties with the page that is actually about them.
+_BODY_MATCH_BONUS = 0.35
 
 
 def _compute_idf(G: nx.Graph, terms: list[str]) -> dict[str, float]:
@@ -360,13 +371,30 @@ def _node_rationale_text(data: dict) -> str:
     return _strip_diacritics(str(raw)).lower()
 
 
+def _node_body_text(data: dict) -> str:
+    """The node's `body` attribute normalized like a label (#3313).
+
+    The documented convention for free text a non-extraction emitter (a
+    Confluence page, a Jira issue) has nowhere else to put. Same shape as
+    `_node_rationale_text`: a list is joined, missing or empty is "".
+    """
+    raw = data.get("body")
+    if not raw:
+        return ""
+    if isinstance(raw, (list, tuple)):
+        raw = " ".join(str(part) for part in raw if part)
+    return _strip_diacritics(str(raw)).lower()
+
+
 def _node_search_text(data: dict, nid: str) -> str:
     """Concatenate every field _score_nodes / _find_node match a query against, so
     one trigram index over this text is a complete candidate generator for both.
 
     - `rationale` (normalized via `_node_rationale_text`) feeds _score_nodes'
       rationale tier (#2293); appended last, and only when present, so every
       other field position is unchanged.
+    - `body` (normalized via `_node_body_text`) feeds _score_nodes' body tier
+      (#3313); same append-only-when-present treatment as rationale.
 
     - `norm_label` and `source_file` feed _score_nodes' per-term substring tiers.
     - `label_tokens` (the space-joined token form) feeds _find_node's
@@ -401,6 +429,9 @@ class 0 and therefore survive the combining-character filter. The field is
     rationale = _node_rationale_text(data)
     if rationale:
         fields += (rationale,)
+    body = _node_body_text(data)
+    if body:
+        fields += (body,)
     return "\x00".join(fields)
 
 
@@ -535,6 +566,18 @@ def _score_query(
     trigram candidate set (needles `norm_terms + [joined]`) is a superset of
     each per-token `[t]` candidate set, so iterating combined candidates
     discovers every non-zero singleton-score node for every term.
+
+    `label`, `source_file`, `rationale` (#2293), and `body` (#3313) are the
+    only attributes matched against. An emitter with nowhere else to put free
+    text (a Confluence page, a Jira issue) can use `body`; anything stored
+    under another attribute name is still invisible to seeding, so the node
+    can never become a seed no matter how relevant its content is - the data
+    is in the graph and the ranker is fine, the node is just never visited,
+    which makes this easy to hit and hard to diagnose. `body` is weighted
+    well below `rationale` rather than sharing its tier: unweighted body
+    matching was measured to make retrieval worse, since a few hundred
+    characters of prose matches many query terms and a page that merely
+    mentions the words ties with the page that is actually about them.
     """
     scored: list[tuple[float, str]] = []
     # Dedupe tokens, order-preserving (as _pick_seeds already does): a repeated
@@ -578,6 +621,7 @@ def _score_query(
         label_tokens = " ".join(_search_tokens(data.get("label") or ""))
         source = (data.get("source_file") or "").lower()
         rationale = _node_rationale_text(data)
+        body = _node_body_text(data)
         # `nid_lower` is needed both by the full-query tier (`if joined`) and by
         # the per-token singleton tier (joined-singlet exact-match check). When
         # neither runs (`joined` empty AND not collecting seeds) skip the call;
@@ -643,6 +687,15 @@ def _score_query(
             if rationale and t in rationale:
                 rationale_value = _RATIONALE_MATCH_BONUS * w
                 score += rationale_value
+            # Body tier (#3313): recall for content a non-extraction emitter
+            # stored on the node with nowhere else to put it. Same shape as
+            # the rationale tier - adds to the score, not to `matched` - at a
+            # lower weight, since unlike a curated rationale, raw body text is
+            # noisy prose where many terms coincidentally appear.
+            body_value = 0.0
+            if body and t in body:
+                body_value = _BODY_MATCH_BONUS * w
+                score += body_value
             tiered += tier_value
             if collect_per_term_seeds and best_by_term is not None:
                 # Singleton score for [t] on this node, mirroring
@@ -661,7 +714,7 @@ def _score_query(
                     singleton = _PREFIX_MATCH_BONUS * 10 * w
                 else:
                     singleton = 0.0
-                singleton += tier_value + substr_value + source_value + rationale_value
+                singleton += tier_value + substr_value + source_value + rationale_value + body_value
                 if singleton > 0:
                     # Tie-break key mirrors the legacy sort+max(degree):
                     # (-singleton, -degree, label_len, nid) — the minimum

tests/test_serve.py

@@ -1827,3 +1827,66 @@ def test_query_graph_text_seeds_the_node_whose_rationale_answers_a_why_question(
     )
     header = text.split("\n\n", 1)[0]
     assert "FAB visibility rule" in header, header
+
+
+# --- body attribute scoring (#3313) ---
+
+def test_score_nodes_reads_body_when_label_does_not_match():
+    """The #3313 shape: a non-extraction emitter (a wiki page, a ticket) puts
+    its content on `body` since there is nowhere else to put it, and the
+    label alone shares no query term."""
+    G = nx.Graph()
+    G.add_node(
+        "page", label="Masterlist/Certificate controls", source_file="confluence/page1.md",
+        body="The CSCA Master List defines certificate authority trust anchors.",
+    )
+    G.add_node("ticket", label="CSCA rollout ticket", source_file="jira/TICKET-1.md")
+    G.add_node("other", label="Unrelated page", source_file="confluence/other.md",
+               body="Nothing to do with certificates at all.")
+    assert "page" in [nid for _, nid in _score_nodes(G, ["CSCA", "Master", "List"])]
+
+
+def test_score_nodes_body_tier_sits_below_rationale_and_source():
+    """Body is the noisiest signal (raw prose, not curated why-text), so it
+    must rank below both the rationale and source-path tiers, not above."""
+    G = nx.Graph()
+    G.add_node("lbl", label="popover-anchor", source_file="ui/a.py")
+    G.add_node("rat", label="Sheet drag", source_file="ui/b.py", rationale="starts only once the popover is closed")
+    G.add_node("src", label="Thing", source_file="ui/popover/thing.py")
+    G.add_node("bod", label="Notes", source_file="ui/c.py", body="mentions a popover somewhere in the middle")
+    assert [nid for _, nid in _score_nodes(G, ["popover"])] == ["lbl", "rat", "src", "bod"]
+
+
+def test_score_nodes_body_does_not_count_toward_term_coverage():
+    """Like rationale and source, a body hit adds recall but must not restore
+    the coverage-scaled exact tier."""
+    from graphify.serve import _EXACT_MATCH_BONUS
+    G = nx.Graph()
+    G.add_node("a", label="cache", source_file="x.py", body="invalidated on every write")
+    G.add_node("b", label="cache", source_file="y.py")
+    score = {nid: s for s, nid in _score_nodes(G, ["cache", "invalidated"])}
+    assert score["a"] > score["b"]
+    assert score["a"] - score["b"] < _EXACT_MATCH_BONUS * 0.5
+
+
+def test_score_nodes_tolerates_list_valued_body():
+    G = nx.Graph()
+    G.add_node("n", label="X", source_file="x.py", body=["first paragraph", "popover second"])
+    assert [nid for _, nid in _score_nodes(G, ["popover"])] == ["n"]
+
+
+def test_node_search_text_includes_body_so_trigram_prefilter_stays_complete():
+    parts = _node_search_text(
+        {"label": "Foo", "source_file": "a.py", "body": "Describes the Popover in detail"}, "foo"
+    ).split("\x00")
+    assert "describes the popover in detail" in parts
+    # No body: field layout unchanged (the #2467 positions still hold).
+    assert len(_node_search_text({"label": "Foo", "source_file": "a.py"}, "foo").split("\x00")) == 5
+
+
+def test_node_search_text_includes_both_rationale_and_body_when_both_present():
+    parts = _node_search_text(
+        {"label": "Foo", "source_file": "a.py", "rationale": "why text", "body": "body text"}, "foo"
+    ).split("\x00")
+    assert "why text" in parts
+    assert "body text" in parts