cfsa-antigravity 1.0.0

package/template/.agent/skill-library/stack/databases/lancedb/SKILL.md

@@ -0,0 +1,415 @@
+---
+name: lancedb
+description: "Expert LanceDB vector database guide covering embedded serverless setup, table design with Pydantic and Arrow schemas, vector search (IVF-PQ, HNSW-PQ), full-text search, hybrid search with rerankers, scalar indexing, embedding functions, async API, cross-store coordination, and data versioning. Use when building RAG pipelines, semantic search, recommendation engines, or any embedding-based retrieval system with an embedded (no-server) vector store."
+version: 1.0.0
+---
+
+# LanceDB Expert Guide
+
+> Use this skill when designing vector tables, writing search queries, planning indexing strategies, setting up LanceDB connections, coordinating LanceDB with primary stores, or choosing between LanceDB and a managed vector database. This skill targets LanceDB 0.x (Python `lancedb` / TypeScript `@lancedb/lancedb`).
+
+## When to Use This Skill
+
+- Building semantic search with an **embedded** (serverless, no-infra) vector store
+- Implementing RAG (Retrieval-Augmented Generation) pipelines with vector retrieval
+- Hybrid search combining vector similarity + full-text search + metadata filtering
+- Multi-modal retrieval (text, image, audio embeddings) in a single table
+- Projects that need **zero-config vector storage** — no Docker, no managed service
+- Integrating as the `DATABASE_VECTOR` store alongside a primary relational/document database
+
+## When NOT to Use This Skill
+
+- Need a managed, distributed vector cluster → use Qdrant or Pinecone
+- Relational data with JOINs → use PostgreSQL
+- Graph traversals → use Neo4j
+- Time-series analytics → use TimescaleDB
+- Key-value caching → use Redis
+- Full-text search without vectors → use Meilisearch or Elasticsearch
+
+---
+
+## 1. Table Design
+
+### Connection
+
+```python
+import lancedb
+
+# Local embedded (file-based — no server required)
+db = lancedb.connect("./my_lancedb")
+
+# Cloud (LanceDB Cloud)
+db = lancedb.connect(
+    "db://my-database",
+    api_key=os.environ["LANCEDB_API_KEY"],
+    region="us-east-1",
+)
+```
+
+```typescript
+import * as lancedb from "@lancedb/lancedb";
+
+// Local embedded
+const db = await lancedb.connect("./my_lancedb");
+
+// Cloud
+const db = await lancedb.connect("db://my-database", {
+  apiKey: process.env.LANCEDB_API_KEY,
+  region: "us-east-1",
+});
+```
+
+### Schema Definition (Python — Pydantic)
+
+Always define schemas with Pydantic models or PyArrow — never raw dicts.
+
+```python
+from lancedb.pydantic import LanceModel, Vector
+from lancedb.embeddings import get_registry
+
+# Auto-embed with an embedding function
+embeddings = get_registry().get("openai").create(name="text-embedding-3-small")
+
+class Document(LanceModel):
+    text: str = embeddings.SourceField()            # source text
+    vector: Vector(1536) = embeddings.VectorField()  # auto-generated
+    category: str
+    author_id: str    # canonical ID from primary store
+    created_at: str
+    public: bool = True
+
+table = db.create_table("documents", schema=Document, mode="overwrite")
+```
+
+### Schema Definition (Python — PyArrow)
+
+```python
+import pyarrow as pa
+
+schema = pa.schema([
+    ("id", pa.int64()),
+    ("text", pa.string()),
+    ("category", pa.string()),
+    ("vector", pa.list_(pa.float32(), 1536)),
+])
+table = db.create_table("documents", schema=schema)
+```
+
+### Schema Definition (TypeScript — Arrow)
+
+```typescript
+import * as lancedb from "@lancedb/lancedb";
+import "@lancedb/lancedb/embedding/openai";
+import { Utf8 } from "apache-arrow";
+
+const embedFunc = lancedb
+  .getRegistry()
+  .get("openai")
+  ?.create({ model: "text-embedding-3-small" });
+
+const schema = lancedb.embedding.LanceSchema({
+  text: embedFunc.sourceField(new Utf8()),
+  vector: embedFunc.vectorField(),
+});
+
+const table = await db.createEmptyTable("documents", schema, {
+  mode: "overwrite",
+});
+```
+
+### Inserting Data
+
+```python
+# With Pydantic model (auto-embeds via SourceField)
+table.add([
+    Document(text="Architecture overview", category="engineering",
+             author_id="user-uuid-1", created_at="2026-01-15"),
+    Document(text="API design patterns", category="engineering",
+             author_id="user-uuid-2", created_at="2026-01-16"),
+])
+
+# With raw dicts (requires pre-computed vectors)
+table.add([
+    {"text": "Hello world", "vector": [0.1, 0.2, ...], "category": "demo"},
+])
+```
+
+```typescript
+await table.add([
+  { text: "Architecture overview" },
+  { text: "API design patterns" },
+]);
+```
+
+---
+
+## 2. Search Patterns
+
+### Basic Vector Search
+
+```python
+# With embedding function — pass text, auto-embeds
+results = table.search("architecture patterns").limit(10).to_pandas()
+
+# With raw vector
+results = table.search([0.1, 0.2, ...]).limit(10).to_pandas()
+```
+
+```typescript
+const results = await table.search("architecture patterns").limit(10).toArray();
+
+// With raw vector
+const results = await table
+  .query()
+  .nearestTo([0.1, 0.2])
+  .limit(10)
+  .toArrow();
+```
+
+### Filtered Search (Vector + Metadata)
+
+```python
+results = (
+    table.search("deployment strategies")
+    .where("category = 'engineering' AND public = true")
+    .limit(20)
+    .to_pandas()
+)
+```
+
+```typescript
+const results = await table
+  .search("deployment strategies")
+  .where("category = 'engineering' AND public = true")
+  .limit(20)
+  .toArray();
+```
+
+### Full-Text Search (FTS)
+
+Requires creating an FTS index first.
+
+```python
+from lancedb.index import FTS
+
+table.create_fts_index("text")
+
+results = table.search("deployment", query_type="fts").limit(10).to_pandas()
+```
+
+```typescript
+await table.createIndex("text", { config: lancedb.Index.fts() });
+
+const results = await table
+  .search("deployment", { queryType: "fts" })
+  .limit(10)
+  .toArray();
+```
+
+### Hybrid Search (Vector + FTS)
+
+Combines vector similarity and full-text search with reranking for best-of-both retrieval.
+
+```python
+from lancedb.rerankers import RRFReranker, LinearCombinationReranker
+
+# Prerequisite: both vector and FTS indexes must exist
+table.create_index("vector", config=lancedb.index.IvfPq())
+table.create_fts_index("text")
+
+# Reciprocal Rank Fusion
+reranker = RRFReranker()
+results = (
+    table.search("AI architecture", query_type="hybrid")
+    .rerank(reranker)
+    .limit(10)
+    .to_pandas()
+)
+
+# Linear combination (70% vector, 30% FTS)
+reranker = LinearCombinationReranker(weight=0.7)
+results = (
+    table.search("AI architecture", query_type="hybrid")
+    .rerank(reranker)
+    .limit(10)
+    .to_pandas()
+)
+```
+
+### Async API (Python)
+
+```python
+import lancedb
+
+async def search():
+    db = await lancedb.connect_async("./my_lancedb")
+    table = await db.open_table("documents")
+    results = await (
+        table.hybrid_search([0.1, 0.2, ...], "query text")
+        .vector_column("vector")
+        .text_column("text")
+        .limit(10)
+        .to_pandas()
+    )
+    return results
+```
+
+---
+
+## 3. Index & Performance
+
+### Vector Indexes
+
+| Index Type | When to Use | Trade-off |
+|-----------|------------|-----------|
+| `IvfPq` | Large datasets (100K+ rows) | Fast, lower recall |
+| `HnswPq` | Medium datasets, higher recall needed | More memory, better accuracy |
+| `IvfFlat` | Binary vectors, Hamming distance | Exact partitioned search |
+
+```python
+import lancedb
+
+# IVF-PQ (recommended for large datasets)
+table.create_index(
+    "vector",
+    config=lancedb.index.IvfPq(
+        num_partitions=256,
+        num_sub_vectors=96,
+        distance_type="cosine",  # "l2", "cosine", or "dot"
+    ),
+)
+
+# HNSW-PQ (better recall, more memory)
+table.create_index(
+    "vector",
+    config=lancedb.index.HnswPq(
+        distance_type="cosine",
+        num_partitions=1,
+        m=20,               # connections per node
+        ef_construction=300, # build-time search width
+    ),
+    replace=True,
+)
+```
+
+```typescript
+await table.createIndex("vector", {
+  config: lancedb.Index.ivfPq({
+    numPartitions: 256,
+    numSubVectors: 96,
+    distanceType: "cosine",
+  }),
+});
+```
+
+### Scalar Indexes (for filter columns)
+
+```python
+# BTree index for filter fields — avoids full scans
+table.create_index("category", config=lancedb.index.BTree())
+```
+
+> Always create scalar indexes on columns used in `.where()` filters — unindexed filter columns trigger full scans.
+
+### Distance Metrics
+
+| Metric | When to Use |
+|--------|------------|
+| `cosine` | Text embeddings (OpenAI, Cohere, most LLMs) — direction matters |
+| `l2` (Euclidean) | Feature vectors where absolute distance matters |
+| `dot` | Pre-normalised embeddings, maximum inner product |
+
+### Listing Indexes
+
+```python
+for idx in table.list_indices():
+    print(f"Index: {idx['name']} on column {idx['columns']}")
+```
+
+---
+
+## 4. Data Versioning
+
+LanceDB uses the Lance columnar format which provides automatic data versioning:
+
+- Every write (add, update, delete) creates a new version
+- Previous versions are retained and queryable
+- Zero-copy reads — multiple readers can access different versions simultaneously
+- Time-travel queries by version number
+
+> This makes LanceDB well-suited for ML experiment tracking and reproducible pipelines where you need to query historical snapshots of your data.
+
+---
+
+## 5. Cross-Store Coordination
+
+### Canonical ID Rule
+
+The primary store's UUID is the canonical identifier. Store it in LanceDB records:
+
+```python
+table.add([
+    Document(
+        text="Architecture overview",
+        category="engineering",
+        author_id="primary-store-uuid-123",  # canonical ID
+        created_at="2026-01-15",
+    ),
+])
+```
+
+### Sync Pattern
+
+1. **Write to primary store first** → generates canonical ID
+2. **Add to LanceDB** with canonical ID (embedding auto-generated if using SourceField)
+3. On primary store update → update or re-add to LanceDB
+4. On primary store delete → delete from LanceDB by filter
+
+```python
+# Delete by filter
+table.delete('author_id = "primary-store-uuid-123"')
+```
+
+---
+
+## 6. Embedding Functions
+
+LanceDB supports built-in embedding functions that auto-generate vectors on insert and search:
+
+| Provider | Registry Key | Example Model |
+|----------|-------------|---------------|
+| OpenAI | `openai` | `text-embedding-3-small`, `text-embedding-3-large` |
+| Cohere | `cohere` | `embed-english-v3.0` |
+| Hugging Face | `huggingface` | `sentence-transformers/all-MiniLM-L6-v2` |
+| Ollama | `ollama` | `nomic-embed-text` |
+
+```python
+from lancedb.embeddings import get_registry
+
+func = get_registry().get("openai").create(name="text-embedding-3-small")
+# Use func with LanceModel SourceField/VectorField (see Section 1)
+```
+
+> When using embedding functions, you pass text to `.search("query text")` and LanceDB auto-embeds it — no manual vector computation needed.
+
+---
+
+## 7. Security & Deployment
+
+- **Embedded mode** — data is local files; protect with filesystem permissions and encryption-at-rest
+- **LanceDB Cloud** — API key auth via `LANCEDB_API_KEY` environment variable
+- **No network exposure** — embedded mode has no server port; no network attack surface
+- **PII in vectors** — embeddings can leak information; apply the same PII controls as the source text
+- **Backups** — Lance format is file-based; standard file backup/replication strategies apply
+
+---
+
+## 8. Common Anti-Patterns
+
+1. **Using raw dicts for schemas** — always use Pydantic `LanceModel` or PyArrow schemas for type safety
+2. **Missing vector indexes on large tables** — tables with 100K+ rows without an IVF-PQ or HNSW index will be slow
+3. **No scalar indexes on filter columns** — `.where()` filters on unindexed columns cause full scans
+4. **Treating LanceDB as a primary store** — LanceDB is a vector index; keep source-of-truth data in a primary store
+5. **Mixed vector dimensions in one table** — all vectors in a column must have the same dimension
+6. **Skipping the embedding function** — manual embedding computation is error-prone; use built-in embedding functions
+7. **Not using hybrid search for RAG** — pure vector search misses keyword-exact matches; hybrid + reranking gives better retrieval