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(generation): keep knowledge-graph layer names unique (#2725)

Raghav Chamadiya56a5f5bb4cbc
diff --git a/packages/core/src/repowise/core/generation/knowledge_graph.py b/packages/core/src/repowise/core/generation/knowledge_graph.py
index f8c4689e5..ece49b8a8 100644
--- a/packages/core/src/repowise/core/generation/knowledge_graph.py
+++ b/packages/core/src/repowise/core/generation/knowledge_graph.py
@@ -15,6 +15,7 @@ from typing import Any
import structlog
from repowise.core.generation.entry_points import orientation_entry_points
+from repowise.core.generation.layers import layer_key
from repowise.core.ids import file_path_of, kg_file_path_of
logger = structlog.get_logger(__name__)
@@ -182,6 +183,7 @@ async def _enrich_layers(
# position: positional joins silently corrupt every name when a model
# returns batch-relative indices.
by_id = {layer["id"]: layer for layer in enriched if layer.get("id")}
+ taken: list[str] = []
for batch_start in range(0, len(layers), _LAYER_BATCH_SIZE):
batch = layers[batch_start : batch_start + _LAYER_BATCH_SIZE]
@@ -198,7 +200,9 @@ async def _enrich_layers(
"top_files": top_files,
})
- user_prompt = _build_layer_naming_prompt(batch_context, tech_stack, repo_structure)
+ user_prompt = _build_layer_naming_prompt(
+ batch_context, tech_stack, repo_structure, taken_names=taken
+ )
batch_ids = {layer.get("id") for layer in batch}
try:
@@ -227,14 +231,38 @@ async def _enrich_layers(
target["description"] = item["description"]
except Exception as exc:
logger.warning("kg_layer_naming_batch_failed", error=str(exc))
+ taken.extend(layer["name"] for layer in batch)
+ _dedupe_layer_names(enriched)
return enriched
+def _dedupe_layer_names(layers: list[dict]) -> None:
+ """Make every layer name unique, in place.
+
+ Batches are named independently, so two layers can come back with the same
+ name. The first layer in display order keeps it; a later one is qualified
+ with its id slug, which is unique by construction.
+ """
+ seen: set[str] = set()
+ for layer in layers:
+ name = layer["name"]
+ if name.casefold() in seen:
+ qualifier = layer_key(str(layer.get("id") or "")).replace("-", " ")
+ name = f"{layer['name']} ({qualifier})"
+ suffix = 2
+ while name.casefold() in seen:
+ name = f"{layer['name']} ({qualifier} {suffix})"
+ suffix += 1
+ layer["name"] = name
+ seen.add(name.casefold())
+
+
def _build_layer_naming_prompt(
batch_context: list[dict],
tech_stack: list[dict],
repo_structure: Any,
+ taken_names: list[str] | None = None,
) -> str:
tech_names = [t.get("name", "") for t in tech_stack[:10] if t.get("name")]
entry_points = orientation_entry_points(repo_structure, limit=5)
@@ -251,8 +279,14 @@ def _build_layer_naming_prompt(
f"Tech stack: {', '.join(tech_names) if tech_names else 'unknown'}",
f"Entry points: {', '.join(entry_points) if entry_points else 'none detected'}",
"",
- "Communities:",
]
+ if taken_names:
+ lines += [
+ "Names already given to other communities (every name must be unique): "
+ + ", ".join(taken_names),
+ "",
+ ]
+ lines.append("Communities:")
for ctx in batch_context:
lines.append(f"\n--- Community \"{ctx['id']}\" ---")
diff --git a/tests/unit/generation/test_knowledge_graph_enrichment.py b/tests/unit/generation/test_knowledge_graph_enrichment.py
index 884bcf69b..31e752d54 100644
--- a/tests/unit/generation/test_knowledge_graph_enrichment.py
+++ b/tests/unit/generation/test_knowledge_graph_enrichment.py
@@ -157,6 +157,45 @@ class TestEnrichLayers:
assert all(layer["name"] != "Clobbered" for layer in result.layers)
assert result.layers[5]["name"] == "heuristic-5" # untouched by bad batch
+ @pytest.mark.asyncio
+ async def test_names_stay_unique_across_batches(self):
+ # Batches are named independently; a name one batch already used must
+ # be offered as taken to the next and, if reused anyway, qualified.
+ layers = [
+ {"id": f"layer:l{i}", "name": f"heuristic-{i}", "description": "",
+ "nodeIds": [f"file:src/f{i}.py"]}
+ for i in range(6)
+ ]
+ nodes = [
+ {"id": f"file:src/f{i}.py", "type": "file", "filePath": f"src/f{i}.py", "summary": ""}
+ for i in range(6)
+ ]
+ responses = iter([
+ '{"layers": [{"id": "layer:l0", "name": "Shared Widgets", "description": "d"}]}',
+ '{"layers": [{"id": "layer:l5", "name": "shared widgets", "description": "d"}]}',
+ '{"tour": []}',
+ ])
+ prompts: list[str] = []
+
+ async def _side_effect(system, user, **kwargs):
+ prompts.append(user)
+ return SimpleNamespace(content=next(responses), input_tokens=1, output_tokens=1)
+
+ llm = AsyncMock()
+ llm.generate.side_effect = _side_effect
+
+ skeleton = _make_kg_skeleton(layers=layers, nodes=nodes)
+ result = await enrich_knowledge_graph(
+ skeleton, llm, _make_graph_builder(), _make_repo_structure(), []
+ )
+
+ assert "Names already given" not in prompts[0]
+ assert "Shared Widgets" in prompts[1]
+ assert result.layers[0]["name"] == "Shared Widgets"
+ assert result.layers[5]["name"] == "shared widgets (l5)"
+ names = [layer["name"].casefold() for layer in result.layers]
+ assert len(names) == len(set(names))
+
@pytest.mark.asyncio
async def test_unknown_id_is_skipped(self):
llm = _make_llm_client(