<4.12.0
V4.5 (Complex Update Required)
FastGPT V4.5 Update
FastGPT V4.5 introduces PgVector 0.5's HNSW index, which dramatically improves knowledge base search performance — roughly 3x to 10x faster than the IVFFlat index, easily achieving millisecond-level searches across millions of records. The downside is that index building is very slow: on a 4C16G machine with 5 million records, a parallel build took about 48 hours. For detailed parameter configuration, refer to the PgVector official documentation.
The following database operations are required for the upgrade:
PgVector Upgrade: Sealos Deployment
- Open the Database app from the Sealos Desktop.
- Click on the details of the pg database.
- Click Restart in the upper-right corner and wait for it to complete.
- Click "One-Click Connect" on the left sidebar and wait for the Terminal to open.
- Run the following SQL commands in order:
-- Upgrade the extension
ALTER EXTENSION vector UPDATE;
-- Verify the upgrade was successful — the vector extension version should be 0.5.0 (previously 0.4.1)
\dx
-- The following two statements set the memory available to PG during index building. Adjust based on your database specs — a good rule of thumb is 1/4 of total memory.
alter system set maintenance_work_mem = '2400MB';
select pg_reload_conf();
-- Rebuild database indexes and collation
REINDEX DATABASE postgres;
-- Start building the index. This takes a very long time — just close the Terminal by clicking the X in the upper-right corner.
CREATE INDEX CONCURRENTLY vector_index ON modeldata USING hnsw (vector vector_ip_ops) WITH (m = 16, ef_construction = 64);
-- You can reconnect to the Terminal and run the command below. If you see "vector_index" hnsw (vector vector_ip_ops) WITH (m='16', ef_construction='64'), the build is complete (make sure there is no INVALID at the end).
\d modeldata |  |
 |  |
PgVector Upgrade: Docker Compose Deployment
The commands below are based on the provided docker-compose template. If you've changed the database username or password, adjust accordingly.
- Update the PG image version in
docker-compose.ymltoankane/pgvector:v0.5.0orregistry.cn-hangzhou.aliyuncs.com/fastgpt/pgvector:v0.5.0. - Restart the PG container (
docker-compose pull && docker-compose up -d) and wait for it to complete. - Enter the container:
docker exec -it pg bash - Connect to the database:
psql 'postgresql://username:password@localhost:5432/postgres' - Run the following SQL commands:
-- Upgrade the extension
ALTER EXTENSION vector UPDATE;
-- Verify the upgrade was successful — the vector extension version should be 0.5.0 (previously 0.4.2)
\dx
-- The following two statements set the memory available to PG during index building. Adjust based on your database specs — a good rule of thumb is 1/4 of total memory.
alter system set maintenance_work_mem = '2400MB';
select pg_reload_conf();
-- Rebuild database indexes and collation
REINDEX DATABASE postgres;
ALTER DATABASE postgres REFRESH COLLATION VERSION;
-- Start building the index. This takes a very long time — just close the terminal window. Do NOT use ctrl+c to cancel.
CREATE INDEX CONCURRENTLY vector_index ON modeldata USING hnsw (vector vector_ip_ops) WITH (m = 16, ef_construction = 64);
-- You can reconnect to the database and run the command below. If you see "vector_index" hnsw (vector vector_ip_ops) WITH (m='16', ef_construction='64'), the build is complete (make sure there is no INVALID at the end).
\d modeldata
New Features
Fast GPT V4.5
- New - Upgraded PgVector extension with HNSW index, dramatically improving knowledge base search speed.
- New - AI Chat node now includes a "Return AI Content" option, allowing you to prevent AI responses from being sent directly to the browser.
- New - Support for selecting models in the Question Classifier.
- Improved - TextSplitter now uses a recursive splitting approach.
- Improved - Advanced orchestration UX performance.
- Fixed - Share link authentication issue.
Configuration File Changes Required
For the latest configuration, refer to: V4.5 Latest config.json
Edit on GitHub
File Updated