@pentatonic-ai/ai-agent-sdk 0.6.0 → 0.7.0

package/packages/memory-engine/engine/services/l2/Dockerfile

@@ -0,0 +1,41 @@
+FROM python:3.12-slim
+
+WORKDIR /app
+
+# System deps for sentence-transformers + neo4j driver
+RUN apt-get update && apt-get install -y --no-install-recommends \
+        gcc g++ git curl \
+    && rm -rf /var/lib/apt/lists/*
+
+# Reranker = sentence-transformers MiniLM cross-encoder.
+# Torch CPU wheels are fine — reranker is small enough to be CPU-bound.
+RUN pip install --no-cache-dir \
+        fastapi "uvicorn[standard]" httpx requests pydantic \
+        neo4j \
+        "sentence-transformers" \
+        "torch" --extra-index-url https://download.pytorch.org/whl/cpu
+
+COPY l2-hybridrag-proxy.py /app/server.py
+COPY init_databases.py /app/init_databases.py
+
+# Pre-create empty L0 BM25 SQLite + empty workspace
+RUN mkdir -p /data/workspace /data/.pentatonic/memory /data/.cache && \
+    python init_databases.py /data
+
+# L2 proxy expects various data dirs — point them at /data
+ENV PME_WORKSPACE=/data/workspace
+ENV PME_QMD_DB=/data/qmd.sqlite
+ENV PME_MEMORY_DB=/data/workspace.db
+ENV HF_HOME=/data/.cache/huggingface
+
+# OSS-friendly defaults: HyDE off (per request), reranker on
+ENV PME_HYDE_ENABLED=false
+ENV PME_RERANK_ENABLED=true
+ENV PME_NV_EMBED_ENABLED=true
+
+EXPOSE 8031
+
+# Pre-download the MiniLM cross-encoder so cold-start is fast
+RUN python -c "from sentence_transformers import CrossEncoder; CrossEncoder('cross-encoder/ms-marco-MiniLM-L-6-v2', cache_folder='/data/.cache/huggingface')"
+
+CMD ["python", "server.py", "--port", "8031", "--host", "0.0.0.0"]

package/packages/memory-engine/engine/services/l2/init_databases.py

@@ -0,0 +1,81 @@
+"""Initialise empty L0 BM25 SQLite + L4 QMD SQLite at container startup.
+
+The L2 proxy reads from two SQLite files that don't otherwise exist
+in a fresh container — so opening them via sqlite3.connect() fails
+with "unable to open database file." Pre-creating them with the
+schemas the proxy expects lets the proxy come up cleanly and return
+empty results from those layers (instead of crashing) until the shim
+populates them through /index-batch.
+
+Schemas mirror what the production ingester writes.
+"""
+
+import os
+import sqlite3
+import sys
+from pathlib import Path
+
+
+def init_l0_bm25(db_path: Path) -> None:
+    """L0 — workspace BM25 over chunked markdown."""
+    db_path.parent.mkdir(parents=True, exist_ok=True)
+    conn = sqlite3.connect(str(db_path))
+    conn.execute("PRAGMA journal_mode=WAL")
+    conn.execute("""
+        CREATE TABLE IF NOT EXISTS chunks (
+            id INTEGER PRIMARY KEY,
+            path TEXT,
+            text TEXT,
+            file_type TEXT,
+            chunk_index INTEGER,
+            created_at TEXT
+        )
+    """)
+    conn.execute("""
+        CREATE VIRTUAL TABLE IF NOT EXISTS chunks_fts USING fts5(
+            path, text, file_type,
+            content='chunks',
+            content_rowid='id'
+        )
+    """)
+    conn.execute("""
+        CREATE TRIGGER IF NOT EXISTS chunks_ai AFTER INSERT ON chunks BEGIN
+            INSERT INTO chunks_fts(rowid, path, text, file_type)
+            VALUES (new.id, new.path, new.text, new.file_type);
+        END
+    """)
+    conn.commit()
+    conn.close()
+
+
+def init_l4_qmd(db_path: Path) -> None:
+    """L4 — QMD vector store, JSONB embeddings stored alongside text."""
+    db_path.parent.mkdir(parents=True, exist_ok=True)
+    conn = sqlite3.connect(str(db_path))
+    conn.execute("PRAGMA journal_mode=WAL")
+    conn.execute("""
+        CREATE TABLE IF NOT EXISTS chunks (
+            id INTEGER PRIMARY KEY,
+            path TEXT,
+            text TEXT,
+            embedding TEXT,
+            embedding_model TEXT,
+            embedding_dim INTEGER,
+            chunk_index INTEGER,
+            created_at TEXT
+        )
+    """)
+    conn.commit()
+    conn.close()
+
+
+def main() -> int:
+    base = Path(sys.argv[1] if len(sys.argv) > 1 else "/data")
+    init_l0_bm25(base / "workspace.db")
+    init_l4_qmd(base / "qmd.sqlite")
+    print(f"initialised L0 + L4 QMD databases under {base}")
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())