Skip to content

Commit

Permalink
doc: pgvector compatibility
Browse files Browse the repository at this point in the history
Signed-off-by: cutecutecat <junyuchen@tensorchord.ai>
  • Loading branch information
cutecutecat committed Jan 22, 2024
1 parent 68b56ea commit b68ab71
Show file tree
Hide file tree
Showing 2 changed files with 120 additions and 0 deletions.
1 change: 1 addition & 0 deletions .vitepress/config.mts
Original file line number Diff line number Diff line change
Expand Up @@ -67,6 +67,7 @@ export default defineConfig({
{ text: 'Search', link: '/usage/search' },
{ text: 'Monitoring', link: '/usage/monitoring' },
{ text: 'Quantization', link: '/usage/quantization' },
{ text: 'Compatibility', link: '/usage/compatibility' },
]
},
{
Expand Down
119 changes: 119 additions & 0 deletions src/usage/compatibility.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,119 @@
# Pgvector compatibility

Pgvecto.rs is natively compatible with pgvector at:
- Create table like `CREATE TABLE t (val vector(3))`
- Query vectors like `INSERT INTO t (val) VALUES ('[0.6,0.6,0.6]')`

Pgvecto.rs can be configured to be compatible with pgvector at:
- Create vector indexes like `CREATE INDEX ON t USING hnsw (val vector_ip_ops);`
- Query options like `SET ivfflat.probes=10;`

This feature is called `compatibility mode`.
It can be enabled with `SET vectors.pgvector_compatibility=on`.

## Options & Variables

For index `ivfflat` and `hnsw` only the following options and variables are available.

* Default value for `options` is **different from pgvecto.rs original**, which keeps the same as in `pgvector`.
* Default value for `variables` is kept the same as in pgvecto.rs.

Options for `ivfflat`:

| Key | Type | Default | Description |
| ----- | ------- | ------- | ------------------------- |
| lists | integer | `100` | Number of cluster units. |

Variables for `ivfflat`:

| Option | Type | Default | Description |
| ---------------- | ------------------------ | ------- | ----------------------------------------- |
| ivfflat.probes | integer (`[1, 1000000]`) | `10` | Number of lists to scan. |

::: warning
Default value of `ivfflat.probes` is `10` instead of `1` from pgvector.
:::

Options for `hnsw`:

| key | type | default | description |
| --------------- | ------- | ------- | -------------------------------- |
| m | integer | `16` | Maximum degree of the node. |
| ef_construction | integer | `64` | Search extent in construction. |

Variables for `hnsw`:

| Option | Type | Default | Description |
| -------------- | ------------------------ | ------- | ----------------------------------------- |
| hnsw.ef_search | integer (`[1, 65535]`) | `100` | Search scope of HNSW. |

::: warning
Default value for `hnsw.ef_search` is `100` instead of `40` from pgvector.
:::

::: tip
The original syntax of pgvecto.rs is still available even in `compatibility mode`.

You can use it with the `vectors` index like before:
```sql
SET vectors.ivf_nprobe=20;
CREATE INDEX ON items USING vectors (embedding vector_l2_ops)
WITH (options = $$
[indexing.ivf]
quantization.product.ratio = "x16"
$$);
```
:::

## Examples

It's easy to enable compatibility mode and start a vector query.
```sql
DROP TABLE IF EXISTS t;
SET vectors.pgvector_compatibility=on;
CREATE TABLE t (val vector(3));
INSERT INTO t (val) SELECT ARRAY[random(), random(), random()]::real[] FROM generate_series(1, 1000);
CREATE INDEX hnsw_l2_index ON t USING hnsw (val vector_l2_ops);
SELECT COUNT(1) FROM (SELECT 1 FROM t ORDER BY val <-> '[0.5,0.5,0.5]' limit 100) t2;
DROP INDEX hnsw_l2_index;
```

Multiply types of indexes are accepted:
```sql
SET vectors.pgvector_compatibility=on;
-- [hnsw + vector_l2_ops] index with default options
CREATE INDEX hnsw_l2_index ON t USING hnsw (val vector_l2_ops);
-- [hnsw + vector_cosine_ops] index with single ef_construction option
CREATE INDEX hnsw_cosine_index ON t USING hnsw (val vector_cosine_ops) WITH (ef_construction = 80);
-- anonymous [hnsw + vector_ip_ops] with all options
CREATE INDEX ON t USING hnsw (val vector_ip_ops) WITH (ef_construction = 80, m = 12);
-- [ivfflat + vector_l2_ops] index with default options
CREATE INDEX ivfflat_l2_index ON t USING ivfflat (val vector_l2_ops);
-- [ivfflat + vector_ip_ops] index with all options
CREATE INDEX ivfflat_ip_index ON t USING ivfflat (val vector_cosine_ops) WITH (nlist = 80);
-- anonymous [ivf + vector_ip_ops] with all options
CREATE INDEX ON t USING ivfflat (val vector_ip_ops) WITH (lists = 80)
```

## Limitation

For compatibility, we strive to maintain a consistent user experience, but there are still some limitations in two aspects:

- Some features of pgvector.rs are not available in `compatibility mode'.
- Some features of pgvector are different in `compatibility mode

### Inaccessible features of pgvecto.rs
When executing SQL statements in `compatibility mode` without `vectors` index, some features of pgvecto.rs are not accessible:
- `flat` index
- Quantization, including `scalar quantization` and `product quantization`
- Prefilter and vbase

### Difference from pgvector

Pgvecto.rs `compatibility mode` only provide compatibility of **creating index** at the foremost.
Nevertheless, there are still differences from pgvector.

Known problems are not limited to:
- Create `btree` index on vector type data is not supported
- Create vector indexes is asynchronous at pgvecto.rs, instead of synchronous at pgvector
- Default value of `variable` is different from pgvector

0 comments on commit b68ab71

Please sign in to comment.