fix: use full embedding context

Co-authored-by: Codex <noreply@openai.com>

README.md

@@ -240,7 +240,7 @@ ADVISOR_REFRESH_LOCK_TTL_SECONDS=7200
 ADVISOR_REFRESH_EMBEDDING_TIMEOUT_SECONDS=1800
 ADVISOR_EMBEDDING_MODEL_REPO=ggml-org/embeddinggemma-300m-qat-q8_0-GGUF
 ADVISOR_EMBEDDING_MODEL_FILE=embeddinggemma-300m-qat-Q8_0.gguf
-ADVISOR_EMBEDDING_N_CTX=768
+ADVISOR_EMBEDDING_N_CTX=2048
 ADVISOR_ASR_MODEL_ID=nvidia/nemotron-speech-streaming-en-0.6b
 ```
 

hackathon_advisor/llama_embedding.py

@@ -19,7 +19,7 @@ from hackathon_advisor.data import (
 
 TRUE_VALUES = {"1", "true", "yes", "on"}
 FALSE_VALUES = {"0", "false", "no", "off"}
-DEFAULT_N_CTX = 768
+DEFAULT_N_CTX = 2048
 
 
 class LlamaCppEmbedder:

scripts/build_project_index.py

@@ -73,7 +73,12 @@ def build_payload(
     data = json.loads(project_path.read_text(encoding="utf-8"))
     projects = [Project.from_dict(item) for item in data["projects"]]
     print(f"loaded {len(projects)} projects from {project_path}", flush=True)
-    reusable_vectors = load_reusable_vectors(reuse_index_path)
+    reusable_vectors = load_reusable_vectors(
+        reuse_index_path,
+        model_repo=model_repo,
+        model_file=model_file,
+        n_ctx=n_ctx,
+    )
     if reusable_vectors:
         print(f"loaded {len(reusable_vectors)} reusable vectors from {reuse_index_path}", flush=True)
     print(
@@ -128,10 +133,41 @@ def build_payload(
     )
 
 
-def load_reusable_vectors(reuse_index_path: Path | None) -> dict[tuple[str, str], list[float]]:
+def load_reusable_vectors(
+    reuse_index_path: Path | None,
+    *,
+    model_repo: str,
+    model_file: str,
+    n_ctx: int,
+) -> dict[tuple[str, str], list[float]]:
     if reuse_index_path is None:
         return {}
     payload = json.loads(reuse_index_path.read_text(encoding="utf-8"))
+    embedding = payload.get("embedding")
+    if not isinstance(embedding, dict):
+        print(f"skipping reusable vectors from {reuse_index_path}: missing embedding metadata", flush=True)
+        return {}
+    expected = {
+        "model_repo": model_repo,
+        "model_file": model_file,
+        "n_ctx": n_ctx,
+    }
+    try:
+        actual_n_ctx = int(embedding.get("n_ctx") or 0)
+    except (TypeError, ValueError):
+        actual_n_ctx = 0
+    actual = {
+        "model_repo": str(embedding.get("model_repo") or ""),
+        "model_file": str(embedding.get("model_file") or ""),
+        "n_ctx": actual_n_ctx,
+    }
+    if actual != expected:
+        print(
+            f"skipping reusable vectors from {reuse_index_path}: "
+            f"embedding config changed from {actual} to {expected}",
+            flush=True,
+        )
+        return {}
     documents = payload.get("documents")
     if not isinstance(documents, list):
         return {}

tests/test_build_project_index.py

@@ -33,7 +33,16 @@ def test_build_project_index_reuses_matching_digest_vectors(monkeypatch, tmp_pat
         encoding="utf-8",
     )
     reuse_path.write_text(
-        json.dumps({"documents": [{"project_id": project.id, "text_digest": digest, "vector": [1.0, 0.0, 0.0]}]}),
+        json.dumps(
+            {
+                "embedding": {
+                    "model_repo": "test/repo",
+                    "model_file": "model.gguf",
+                    "n_ctx": build_project_index.DEFAULT_N_CTX,
+                },
+                "documents": [{"project_id": project.id, "text_digest": digest, "vector": [1.0, 0.0, 0.0]}],
+            }
+        ),
         encoding="utf-8",
     )
 
@@ -55,3 +64,66 @@ def test_build_project_index_reuses_matching_digest_vectors(monkeypatch, tmp_pat
     assert payload["documents"][0]["project_id"] == project.id
     assert payload["documents"][0]["text_digest"] == digest
     assert payload["documents"][0]["vector"] == [1.0, 0.0, 0.0]
+
+
+def test_build_project_index_rejects_vectors_when_embedding_config_changes(monkeypatch, tmp_path: Path) -> None:
+    project_row = {
+        "id": "build-small-hackathon/rebuilt-project",
+        "title": "Rebuilt Project",
+        "summary": "compact local model demo",
+        "tags": ["gradio"],
+        "models": [],
+        "datasets": [],
+        "likes": 0,
+        "sdk": "gradio",
+        "license": "",
+        "created_at": "",
+        "last_modified": "",
+        "host": "",
+        "url": "https://example.test",
+    }
+    project = Project.from_dict(project_row)
+    digest = sha256(project.searchable_text.encode("utf-8")).hexdigest()
+    project_path = tmp_path / "projects.json"
+    reuse_path = tmp_path / "reuse.json"
+    project_path.write_text(
+        json.dumps({"generated_at": "2026-06-08T00:00:00+00:00", "source": "test", "projects": [project_row]}),
+        encoding="utf-8",
+    )
+    reuse_path.write_text(
+        json.dumps(
+            {
+                "embedding": {
+                    "model_repo": "test/repo",
+                    "model_file": "model.gguf",
+                    "n_ctx": 768,
+                },
+                "documents": [{"project_id": project.id, "text_digest": digest, "vector": [1.0, 0.0, 0.0]}],
+            }
+        ),
+        encoding="utf-8",
+    )
+
+    class FakeEmbedder:
+        def __init__(self, **kwargs) -> None:
+            assert kwargs["n_ctx"] == 2048
+
+        def embed(self, _text: str) -> list[float]:
+            return [0.0, 1.0, 0.0]
+
+    monkeypatch.setattr(build_project_index, "LlamaCppEmbedder", FakeEmbedder)
+
+    payload = build_project_index.build_payload(
+        project_path,
+        model_repo="test/repo",
+        model_file="model.gguf",
+        n_ctx=2048,
+        build_source="test",
+        builder="test",
+        reuse_index_path=reuse_path,
+    )
+
+    assert payload["document_count"] == 1
+    assert payload["documents"][0]["project_id"] == project.id
+    assert payload["documents"][0]["text_digest"] == digest
+    assert payload["documents"][0]["vector"] == [0.0, 1.0, 0.0]