v0.31.1

PromptingTools v0.31.1

Diff since v0.31.0

Updated

• Improved the implementation of RAGTools.unpack_bits to be faster with fewer allocations.

Commits

Merged pull requests:

• Improve unpack_bits (#165) (@svilupp)

v0.31.0

PromptingTools v0.31.0

Diff since v0.30.0

Breaking Changes

• The return type of RAGTools.find_tags(::NoTagger,...) is now ::Nothing instead of CandidateChunks/MultiCandidateChunks with all documents.
• Base.getindex(::MultiIndex, ::MultiCandidateChunks) now always returns sorted chunks for consistency with the behavior of other getindex methods on *Chunks.

Updated

• Cosine similarity search now uses partialsortperm for better performance on large datasets.
• Skip unnecessary work when the tagging functionality in the RAG pipeline is disabled (find_tags with NoTagger always returns nothing which improves the compiled code).
• Changed the default behavior of getindex(::MultiIndex, ::MultiCandidateChunks) to always return sorted chunks for consistency with other similar functions. Note that you should always use re-rankering anyway (see FlashRank.jl).

Commits

Merged pull requests:

• Update RAG performance (Breaking changes) (#164) (@svilupp)

v0.30.0

PromptingTools v0.30.0

Diff since v0.29.0

Fixed

• Fixed a bug on Julia 1.11 beta by adding REPL stdlib as a direct dependency.
• Fixed too restrictive argument types for RAGTools.build_tags method.

Commits

Merged pull requests:

• Update deps for 1.11 (REPL) (#163) (@svilupp)

v0.29.0

PromptingTools v0.29.0

Diff since v0.28.0

Added

• Added a package extension for FlashRank.jl to support local ranking models. See ?RT.FlashRanker for more information or examples/RAG_with_FlashRank.jl for a quick example.

Commits

Merged pull requests:

• Add FlashRank.jl package extension (#162) (@svilupp)

Closed issues:

• Allow changing OpenAI key after first load (#161)

v0.28.0

PromptingTools v0.28.0

Diff since v0.27.0

### Added

• Added Mistral coding-oriented Codestral to the model registry, aliased as codestral or mistralc. It's very fast, performant and much cheaper than similar models.

### Commits

Merged pull requests:

• Add Mistral Codestral (#160) (@svilupp)

v0.27.0

PromptingTools v0.27.0

Diff since v0.26.2

Added

• Added a keyword-based search similarity to RAGTools to serve both for baseline evaluation and for advanced performance (by having a hybrid index with both embeddings and BM25). See ?RT.KeywordsIndexer and ?RT.BM25Similarity for more information, to build use build_index(KeywordsIndexer(), texts) or convert an existing embeddings-based index ChunkKeywordsIndex(index).

Updated

• For naming consistency, ChunkIndex in RAGTools has been renamed to ChunkEmbeddingsIndex (with an alias ChunkIndex for backwards compatibility). There are now two main index types: ChunkEmbeddingsIndex and ChunkKeywordsIndex (=BM25), which can be combined into a MultiIndex to serve as a hybrid index.

Commits

Merged pull requests:

• Add BM25 Index (#157) (@svilupp)
• Implement E2E Hybrid retrieval (BM25 + embeddings) (#159) (@svilupp)

Closed issues:

• [FR] RAG: Add HybridChunkIndex with BM25 support (#119)

v0.26.2

PromptingTools v0.26.2

Diff since v0.26.1

Fixed

• Fixed a rare bug where prompt templates created on MacOS will come with metadata that breaks the prompt loader. From now on, it ignores any dotfiles (hidden files starting with ".").

Commits

Merged pull requests:

• Template file parsing fix (#156) (@svilupp)

v0.26.1

PromptingTools v0.26.1

Diff since v0.26.0

Fixed

• Fixed a bug where utility length_longest_common_subsequence was not working with complex Unicode characters

Commits

Merged pull requests:

• Fix LCS utility (#155) (@svilupp)

v0.26.0

PromptingTools v0.26.0

Diff since v0.25.0

BREAKING CHANGES

• Added new field meta to TracerMessage and TracerMessageLike to hold metadata in a simply dictionary. Change is backward-compatible.
• Changed behaviour of aitemplates(name::Symbol) to look for the exact match on the template name, not just a partial match. This is a breaking change for the aitemplates function only. Motivation is that having multiple matches could have introduced subtle bugs when looking up valid placeholders for a template.

Added

• Improved support for aiclassify with OpenAI models (you can now encode upto 40 choices).
• Added a template for routing questions :QuestionRouter (to be used with aiclassify)
• Improved tracing by TracerSchema to automatically capture crucial metadata such as any LLM API kwargs (api_kwargs), use of prompt templates and its version. Information is captured in meta(tracer) dictionary. See ?TracerSchema for more information.
• New tracing schema SaverSchema allows to automatically serialize all conversations. It can be composed with other tracing schemas, eg, TracerSchema to automatically capture necessary metadata and serialize. See ?SaverSchema for more information.
• Updated options for Binary embeddings (refer to release v0.18 for motivation). Adds utility functions pack_bits and unpack_bits to move between binary and UInt64 representations of embeddings. RAGTools adds the corresponding BitPackedBatchEmbedder and BitPackedCosineSimilarity for fast retrieval on these Bool<->UInt64 embeddings (credit to domluna's tinyRAG).

Fixed

• Fixed a bug where aiclassify would not work when returning the full conversation for choices with extra descriptions

Commits

Merged pull requests:

• Improvements to aiclassify and aitemplates (#150) (@svilupp)
• Improve tracer schema / automated logging (#151) (@svilupp)
• Add BitPacked embeddings for RAG retrieval (#152) (@svilupp)
• add tracer kwargs (#153) (@svilupp)

Closed issues:

• [FR] Improve performance of Bool embeddings (#144)

v0.25.0

PromptingTools v0.25.0

Diff since v0.24.0

Added

• Added model registry record for the latest OpenAI GPT4 Omni model (gpt4o) - it's as good as GPT4, faster and cheaper.

Commits

Merged pull requests:

• Add GPT4-Omni (#149) (@svilupp)