Skip to content
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension


Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
2 changes: 1 addition & 1 deletion .github/workflows/release.yml
Original file line number Diff line number Diff line change
Expand Up @@ -48,7 +48,7 @@ jobs:

github-release:
runs-on: ubuntu-latest
needs: [pypi, images]
needs: [images] # PyPI publishes on its own; a missing PyPI setup must not block the release
permissions:
contents: write
steps:
Expand Down
26 changes: 13 additions & 13 deletions ROADMAP.md
Original file line number Diff line number Diff line change
Expand Up @@ -6,27 +6,27 @@ how do I get there in Postgres? Items move from Next to Now when someone is work
## Now

- **Dimension truncation.** Matryoshka models (OpenAI text-embedding-3, Nomic, others) can be cut to
fewer dimensions. Compare "fewer dimensions" with "fewer bits", and both together.
fewer dimensions. Compare "fewer dimensions" with "fewer bits", and both together. ([#3](https://github.com/ahmad-codex/shrinkvec/issues/3))
- **Filtered search.** Recall under a `WHERE` clause, with and without pgvector's iterative index
scans, since filters are where approximate indexes lose the most.
- **IVFFlat** on `vector` and `halfvec` in the Postgres runner.
scans, since filters are where approximate indexes lose the most. ([#4](https://github.com/ahmad-codex/shrinkvec/issues/4))
- **IVFFlat** on `vector` and `halfvec` in the Postgres runner. ([#5](https://github.com/ahmad-codex/shrinkvec/issues/5))
- **More extensions when installed:** pgvectorscale (DiskANN with statistical binary quantization)
and pg_turbovec.
- **A text-embedding-3 dataset** with a clear license, next to ada-002, GloVe and SIFT.
and pg_turbovec. ([#6](https://github.com/ahmad-codex/shrinkvec/issues/6), [#7](https://github.com/ahmad-codex/shrinkvec/issues/7))
- **A text-embedding-3 dataset** with a clear license, next to ada-002, GloVe and SIFT. ([#8](https://github.com/ahmad-codex/shrinkvec/issues/8))

## Next

- **Throughput under load:** N concurrent clients, not only single-query latency.
- **Product quantization (PQ)** in the lab, as the classic baseline the papers compare against.
- **RaBitQ with many centroids** and a 4-bit query, to match VectorChord's exact setup in the lab.
- **Cache the ground truth** between runs on the same vectors.
- **A Python API page**, for using the lab from a notebook.
- **Throughput under load:** N concurrent clients, not only single-query latency. ([#9](https://github.com/ahmad-codex/shrinkvec/issues/9))
- **Product quantization (PQ)** in the lab, as the classic baseline the papers compare against. ([#10](https://github.com/ahmad-codex/shrinkvec/issues/10))
- **RaBitQ with many centroids** and a 4-bit query, to match VectorChord's exact setup in the lab. ([#11](https://github.com/ahmad-codex/shrinkvec/issues/11))
- **Cache the ground truth** between runs on the same vectors. ([#12](https://github.com/ahmad-codex/shrinkvec/issues/12))
- **A Python API page**, for using the lab from a notebook. ([#13](https://github.com/ahmad-codex/shrinkvec/issues/13))

## Later

- **Faster lab scans** with integer kernels instead of decoding to floats.
- **Memory view:** the size of each option at 1M and 100M vectors.
- **Compare reports** from several runs (models, datasets) in one page.
- **Faster lab scans** with integer kernels instead of decoding to floats. ([#14](https://github.com/ahmad-codex/shrinkvec/issues/14))
- **Memory view:** the size of each option at 1M and 100M vectors. ([#15](https://github.com/ahmad-codex/shrinkvec/issues/15))
- **Compare reports** from several runs (models, datasets) in one page. ([#16](https://github.com/ahmad-codex/shrinkvec/issues/16))

## Not planned

Expand Down
Loading