Turn pgvector planning into repo-native ingestion templates

Constraint: The user needs concrete downstream data handling guidance now, and future vector retrieval work should not start from abstract docs alone
Rejected: Leave pgvector support at prose-only guidance | Delays integration by forcing later sessions to reinvent schema and export bridges
Confidence: high
Scope-risk: narrow
Directive: Keep schema/export templates aligned with actual manifest semantics before adding live database loaders
Tested: /usr/local/miniconda3/bin/python -m py_compile acr-engine/scripts/export_manifest_to_pgvector_json.py; /usr/local/miniconda3/bin/python acr-engine/scripts/export_manifest_to_pgvector_json.py --data acr-engine/data/synthetic_v2 --split test --source-dataset synthetic_v2 --output acr-engine/reports/pgvector_manifest_export_test.json
Not-tested: Live PostgreSQL/pgvector ingestion remains pending a real database target

+#!/usr/bin/env python3
+"""Export project manifests into a pgvector-friendly JSON payload.
+
+This does not require PostgreSQL at runtime. It prepares normalized rows so a
+future loader can bulk ingest them into Postgres/pgvector safely.
+"""
+
+from __future__ import annotations
+
+import argparse
+import json
+from pathlib import Path
+
+
+def load_json(path: Path):
+    return json.loads(path.read_text())
+
+
+def main():
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--data", required=True, help="manifest directory")
+    parser.add_argument("--output", required=True)
+    parser.add_argument("--split", default="train")
+    parser.add_argument("--source-dataset", default="unknown")
+    args = parser.parse_args()
+
+    data_dir = Path(args.data)
+    catalog = load_json(data_dir / "catalog.json")
+    split_rows = load_json(data_dir / f"{args.split}.json")
+
+    songs = {}
+    references = []
+    segments = []
+
+    for row in catalog:
+        song_id = row["song_id"]
+        songs.setdefault(song_id, {
+            "song_id": song_id,
+            "title": song_id,
+            "artist": None,
+            "version_id": None,
+            "source_dataset": row.get("source_dataset", args.source_dataset),
+            "license": None,
+        })
+        if row.get("type") == "reference":
+            references.append({
+                "song_id": song_id,
+                "audio_uri": row["audio_path"],
+                "duration_sec": row["duration"],
+                "sample_rate": 16000,
+            })
+
+    for row in split_rows:
+        if row.get("type") == "reference":
+            continue
+        song_id = row["song_id"]
+        songs.setdefault(song_id, {
+            "song_id": song_id,
+            "title": song_id,
+            "artist": None,
+            "version_id": None,
+            "source_dataset": row.get("source_dataset", args.source_dataset),
+            "license": None,
+        })
+        segments.append({
+            "song_id": song_id,
+            "audio_uri": row["audio_path"],
+            "offset_sec": row.get("offset", 0.0),
+            "duration_sec": row["duration"],
+            "split": args.split,
+            "type": row.get("type", "unknown"),
+            "segment_type": row.get("segment_type"),
+            "source_dataset": row.get("source_dataset", args.source_dataset),
+        })
+
+    payload = {
+        "songs": list(songs.values()),
+        "references": references,
+        "segments": segments,
+    }
+
+    out = Path(args.output)
+    out.parent.mkdir(parents=True, exist_ok=True)
+    out.write_text(json.dumps(payload, indent=2, ensure_ascii=False))
+    print(json.dumps({
+        "status": "ok",
+        "output": str(out.resolve()),
+        "songs": len(payload["songs"]),
+        "references": len(payload["references"]),
+        "segments": len(payload["segments"]),
+    }, indent=2, ensure_ascii=False))
+
+
+if __name__ == "__main__":
+    main()

+CREATE EXTENSION IF NOT EXISTS vector;
+
+CREATE TABLE IF NOT EXISTS songs (
+    song_id TEXT PRIMARY KEY,
+    title TEXT,
+    artist TEXT,
+    version_id TEXT,
+    source_dataset TEXT,
+    license TEXT,
+    created_at TIMESTAMPTZ DEFAULT NOW()
+);
+
+CREATE TABLE IF NOT EXISTS references (
+    reference_id BIGSERIAL PRIMARY KEY,
+    song_id TEXT NOT NULL REFERENCES songs(song_id) ON DELETE CASCADE,
+    audio_uri TEXT NOT NULL,
+    duration_sec DOUBLE PRECISION NOT NULL,
+    sample_rate INTEGER DEFAULT 16000,
+    created_at TIMESTAMPTZ DEFAULT NOW()
+);
+
+CREATE TABLE IF NOT EXISTS segments (
+    segment_id BIGSERIAL PRIMARY KEY,
+    song_id TEXT NOT NULL REFERENCES songs(song_id) ON DELETE CASCADE,
+    audio_uri TEXT NOT NULL,
+    offset_sec DOUBLE PRECISION DEFAULT 0,
+    duration_sec DOUBLE PRECISION NOT NULL,
+    split TEXT,
+    type TEXT NOT NULL,
+    segment_type TEXT,
+    source_dataset TEXT,
+    created_at TIMESTAMPTZ DEFAULT NOW()
+);
+
+CREATE TABLE IF NOT EXISTS reference_embeddings (
+    embedding_id BIGSERIAL PRIMARY KEY,
+    reference_id BIGINT NOT NULL REFERENCES references(reference_id) ON DELETE CASCADE,
+    song_id TEXT NOT NULL REFERENCES songs(song_id) ON DELETE CASCADE,
+    embedding vector(192) NOT NULL,
+    model_version TEXT NOT NULL,
+    data_version TEXT,
+    created_at TIMESTAMPTZ DEFAULT NOW()
+);
+
+CREATE TABLE IF NOT EXISTS query_embeddings (
+    embedding_id BIGSERIAL PRIMARY KEY,
+    segment_id BIGINT NOT NULL REFERENCES segments(segment_id) ON DELETE CASCADE,
+    song_id TEXT NOT NULL REFERENCES songs(song_id) ON DELETE CASCADE,
+    embedding vector(192) NOT NULL,
+    model_version TEXT NOT NULL,
+    data_version TEXT,
+    created_at TIMESTAMPTZ DEFAULT NOW()
+);
+
+CREATE INDEX IF NOT EXISTS idx_segments_song_id ON segments(song_id);
+CREATE INDEX IF NOT EXISTS idx_references_song_id ON references(song_id);
+CREATE INDEX IF NOT EXISTS idx_reference_embeddings_song_id ON reference_embeddings(song_id);
+CREATE INDEX IF NOT EXISTS idx_query_embeddings_song_id ON query_embeddings(song_id);
+
+CREATE INDEX IF NOT EXISTS idx_reference_embeddings_vector_cosine
+ON reference_embeddings USING ivfflat (embedding vector_cosine_ops)
+WITH (lists = 100);
+
+CREATE INDEX IF NOT EXISTS idx_query_embeddings_vector_cosine
+ON query_embeddings USING ivfflat (embedding vector_cosine_ops)
+WITH (lists = 100);

@@ -234,6 +234,26 @@
 
 
 
 
 
 
+
+### Stage: pgvector 落库模板
+
+完成项：
+- 新增 [acr-engine/sql/pgvector_schema.sql](../acr-engine/sql/pgvector_schema.sql)
+- 新增 [acr-engine/scripts/export_manifest_to_pgvector_json.py](../acr-engine/scripts/export_manifest_to_pgvector_json.py)
+- 在 [docs/training-data-and-pgvector-guide.md](./training-data-and-pgvector-guide.md) 中补充可执行模板说明
+
+验证结果：
+- `/usr/local/miniconda3/bin/python -m py_compile scripts/export_manifest_to_pgvector_json.py` 成功
+- `/usr/local/miniconda3/bin/python scripts/export_manifest_to_pgvector_json.py --data data/synthetic_v2 --split test --source-dataset synthetic_v2 --output reports/pgvector_manifest_export_test.json` 成功
+- 当前导出结果：
+  - `songs=24`
+  - `references=24`
+  - `segments=20`
+
+结论：
+- pgvector 方向现在不仅有概念文档，还有可直接复用的 schema 和 manifest 导出桥接脚本
+- 后续接 PostgreSQL 时返工成本会显著降低
+
 ### Stage: FMA 下载自动守护
 ### Stage: FMA 下载自动守护
 
 
 完成项：
 完成项：

@@ -508,6 +508,37 @@ val.json
 - [dataset-sources-and-licensing.md](./dataset-sources-and-licensing.md)
 - [dataset-sources-and-licensing.md](./dataset-sources-and-licensing.md)
 - [session-handoff.md](./session-handoff.md)
 - [session-handoff.md](./session-handoff.md)
 
 
+
+## 12. 可直接落地的 pgvector 模板
+
+仓库里现在已经补了两个可直接参考的模板：
+
+- SQL schema: [acr-engine/sql/pgvector_schema.sql](../acr-engine/sql/pgvector_schema.sql)
+- manifest 导出桥接脚本: [acr-engine/scripts/export_manifest_to_pgvector_json.py](../acr-engine/scripts/export_manifest_to_pgvector_json.py)
+
+### 导出示例
+
+```bash
+cd acr-engine
+/usr/local/miniconda3/bin/python scripts/export_manifest_to_pgvector_json.py \
+  --data data/synthetic_v2 \
+  --split test \
+  --source-dataset synthetic_v2 \
+  --output reports/pgvector_manifest_export_test.json
+```
+
+### 当前已验证结果
+
+- `songs=24`
+- `references=24`
+- `segments=20`
+
+这一步还不会直接写 PostgreSQL，作用是：
+
+1. 先把项目现有 manifest 规范转换成 pgvector-friendly 结构化 JSON
+2. 后续你们可以再用 bulk insert / COPY / ETL 把这些行落到 PostgreSQL
+3. embedding 生成后再写入 `vector(192)` 列
+
 ## Sources
 ## Sources
 
 
 - Current code behavior from:
 - Current code behavior from: