Skip to content

Changelog

[v0.4.4] - 2025-06-20

Fixed

  • Changed the allowed base branch names in the GHA guard-branches.yml: doc -> docs
  • Fixed kajson dependency (see kajson v0.1.6 changelog)

Cursor rules

  • Added Cursor rules for coding best practices and standards (including linting methods). Added TDD (Test Driven Development) rule on demand.
  • Various changes

Documentation

  • Added documentation for referencing images in PipeLLM.
  • Fixed typos

Refactor

  • Removed the images field from PipeLLM - images can now be referenced directly in the inputs
  • Moved the list-pipes CLI function to the PipeLibrary class.

[v0.4.3] - 2025-06-19

Fixed

  • Removed deprecated Gemini 1.5 models: Removed gemini-1.5-flash and gemini-1.5-pro from the VertexAI integration as they are no longer supported
  • Fixed multiple import statements across the codebase

Documentation

  • Enhanced MkDocs search: Added search functionality to the documentation site
  • Proofreading improvements: Fixed various typos and improved clarity across documentation

Refactor

  • Mini refactor: changed kajson dependency to kajson==0.1.5 (instead of >=) to tolerate temporary breaking changes from kajson

[v0.4.2] - 2025-06-17

  • Fixed the inheritance config manager method (Undocumented feature, soon to be removed)
  • Fixed the deploy-doc.yml GitHub Action
  • Grouped the mkdocs dependencies in a single group docs in the pyproject.toml file

[v0.4.1] - 2025-06-16

  • Changed discord link to the new one: https://go.pipelex.com/discord
  • Added hello-world example in the cookbook-examples of the documentation.

[v0.4.0] - 2025-06-16

Highlight: Complete documentation overhaul

  • MkDocs setup for static web docs generation
    • Material for MkDocs theme, custom styling and navigation
    • Other plugins: meta-manager, glightbox
    • GitHub Pages deployment, mapped to docs.pipelex.com
    • Added GHA workflows for documentation deployment and validation
  • Added to docs:
    • Manifesto explaining the Pipelex viewpoint
    • The Pipelex Paradigm explaining the fundamentals of Pipelex's solution
    • **Cookbook examples** presented and explained, commented code, some event with mermaid flow charts
    • And plenty of details about using Pipelex and developing for Pipelex, from structured generation to PipeOperators (LLM, Image generation, OCR…) to PipeControllers (Sequence, Parallel, Batch, Condition…), workflow optimization, workflow static validation and dry run… there's still work to do, but we move fast!
  • Also a major update of Cursor rules

Tooling Improvements

  • Pipeline tracking: restored visual flowchart generation using Mermaid
  • Enhanced dry run configuration: added more granular control with nb_list_items, nb_ocr_pages, and image_urls
  • New feature flags: better control over pipeline tracking, activity tracking, and reporting
  • Improved OCR configuration: handle image file type for Mistral-OCR, added default_page_views_dpi setting
  • Enhanced LLM configuration: better prompting for structured generation with automatic schema insertion for two-step structuring: generate plain text and then structure via Json
  • Better logging: Enhanced log truncation and display for large objects like image bytes (there are still cases to deal with)

Refactor

Concept system refactoring

  • Improved concept code factory with better domain handling, so you no longer need the native domain prefix for native domains, you can just call them by their names: Text, Image, PDF, Page, Number…
  • Concept refines attribute can now be a string for single refined concepts (the most common case)

Breaking Changes

  • File structure changes: documentation moved from doc/ to docs/
  • Configuration changes: some configuration keys have been renamed or restructured
  • StuffFactory.make_stuff() argument concept_code renamed to concept_str to explicitly support concepts without fully qualified domains (e.g., Text or PDF implicitly native )
  • Some method signatures have been updated

Tests

  • Added Concept refinement validation: TestConceptRefinesValidationFunction and TestConceptPydanticFieldValidation ensure proper concept inheritance and field validation

[v0.3.2] - 2025-06-13

  • Improved automatic insertion of class structure from BaseModel into prompts, based on the PipeLLM's output_concept. New unit test included.
  • The ReportingManager now reports costs for all pipeline IDs when no pipeline_run_id is specified.
  • The make_from_str method from the StuffFactory class now uses Text context by default.

[v0.3.1] - 2025-06-10

Added

  • New pytest marker dry_runnable for tests that can run without inference.
  • Enhanced make targets with dry-run capabilities for improved test coverage:
  • make test-xdist (or make t): Runs all non-inference tests plus inference tests that support dry-runs - fast and resource-efficient
  • make test-inference (or make ti): Runs tests requiring actual inference, with actual inference (slow and costly)
  • Parallel test execution using pytest-xdist (-n auto) enabled for:
  • GitHub Actions workflows
  • Codex test targets

Changed

  • Domain validation is now less restrictive in pipeline TOML: the definition attribute is now Optional

[v0.3.0] - 2025-06-09

Highlights

  • Structured Input Specifications: Pipe inputs are now defined as a dictionary mapping a required variable name to a concept code (required_variable -> concept_code). This replaces the previous single input field and allows for multiple, named inputs, making pipes more powerful and explicit. This is a breaking change.
  • Static Validation for Inference Pipes: You can now catch configuration and input mistakes in your pipelines before running any operations. This static validation checks PipeLLM, PipeOcr, and PipeImgGen. Static validation for controller pipes (PipeSequence, PipeParallel…) will come in a future release.
    • Configure the behavior for different error types using the static_validation_config section in your settings. For each error type, choose to raise, log, or ignore.
  • Dry Run Mode for Zero-Cost Pipeline Validation: A powerful dry-run mode allows you to test entire pipelines without making any actual inference calls. It's fast, costs nothing, works offline, and is perfect for linting and validating pipeline logic.
    • The new dry_run_config lets you control settings, like disabling Jinja2 rendering during a dry run.
    • This feature leverages polyfactory to generate mock Pydantic models for simulated outputs.
    • Error handling for bad inputs during run_pipe has been improved and is fully effective in dry-run mode.
    • One limitation: currently, dry running doesn't work when the pipeline uses a PipeCondition. This will be fixed in a future release.

Added

  • native.Anything Concept: A new flexible native concept that is compatible with any other concept, simplifying pipe definitions where input types can vary.
  • Added dependency on polyfactory for mock Pydantic model generation in dry-run mode.

Changed

  • Refactored Cognitive Workers: The abstraction for LLM, Imgg, and Ocr workers has been elegantly simplified. The old decorator-based approach (..._job_func) has been replaced with a more robust pattern: a public base method now handles pre- and post-execution logic while calling a private abstract method that each worker implements.
  • The b64_image_bytes field in PromptImageBytes was renamed to base_64 for better consistency.

Fixed

  • Resolved a logged error related to the pipe stack when using PipeParallel.
  • The pipe tracker functionality has been restored. It no longer crashes when using nested object attributes (e.g., my_object.attribute) as pipe inputs.

Tests

  • A new pytest command-line option --pipe-run-mode has been added to switch between live and dry runs (default is dry). All pipe tests now respect this mode.
  • Introduced the pipelex_api pytest marker for tests related to the Pipelex API client, separating them from general inference or llm tests.
  • Added a make test-pipelex-api target (shorthand: make ta) to exclusively run these new API client tests.

Removed

  • The llm_job_func.py file and the associated decorators have been removed as part of the cognitive worker refactoring.

[v0.2.14] - 2025-06-06

  • Added a feature flag for the ReportingManager in the config: 
    [pipelex]
[pipelex.feature_config]
is_reporting_enabled = true
  • Moved the reporting config form the cogtconfig to the Pipelex config.

[v0.2.13] - 2025-06-06

  • Added Discord badge on the Readme. Join the community! -> https://go.pipelex.com/discord
  • Added a client for the Pipelex API. Join the waitlist -> https://www.pipelex.com/signup
  • Removed the run_pipe_code function. Replaced by execute_pipeline in pipelex.pipeline.execute.
  • Added llm deck llm_for_img_to_text.
  • Renamed InferenceReportManager to ReportingManager: It can report more than Inference cost. Renamed InferenceReportDelegate to ReportingProtocol.
  • Added an injection of dependency for ReportingManager
  • pipelex cli: fixed some bugs

[v0.2.12] - 2025-06-03

  • pipelex cli: Split pipelex init into 2 separate functions: pipelex init-libraries and pipelex init-config
  • Fixed the inheritance config manager method
  • Rename Mission to Pipeline
  • Enable to start a pipeline and let in run in the background, getting it's run id, but not waiting for the output
  • Makefile: avoid defaulting pytest to verbose. Setup target make test-xdist = Run unit tests with xdist, make it the default for shorthand make t. The old make t is now make tp (test-with-prints)
  • Added mistral-small-3.1 and qwen3:8b
  • Fix template pre-processor: don't try and substitute a dollar numerical like $10 or @25
  • Refactor with less "OpenAI" naming for non-openai stuff that just uses the OpenAI SDK

[v0.2.11] - 2025-06-02

  • HotFix for v0.2.10 👇 regarding the new pipelex/pipelex_init.toml`

[v0.2.10] - 2025-06-02

Highlights

Python Support Expansion - We're no longer tied to Python 3.11! Now supporting Python 3.10, 3.11, 3.12, and 3.13 with full CI coverage across all versions.

Major Model Additions - Claude 4 (Opus & Sonnet), Grok-3, and GPT-4 image generation are now in the house.

Pipeline Base Library update

  • New pipe - extract_page_contents_and_views_from_pdf transferred from cookbook to base library (congrats on the promotion!). This pipe extracts text, linked images, AND page_view images (rendered pages) - it's very useful if you want to use Vision in follow-up pipes

Added

  • Template preprocessor - New @? token prefix for optional variable insertion - if a variable doesn't exist, we gracefully skip it instead of throwing exceptions
  • Claude 4 support - Both Opus and Sonnet variants, available through Anthropic SDK (direct & Bedrock) plus Bedrock SDK. Includes specific max_tokens limit reduction to prevent timeout/streaming issues (temporary workaround)
  • Grok-3 family support - Full support via OpenAI SDK for X.AI's latest models
  • GPT-4 image generation - New gpt-image-1 model through OpenAI SDK, available via PipeImgGen. Currently saves local files (addressing in next release)
  • Gemini update - Added latest gemini-2.5-pro to the lineup
  • Image generation enhancements - Better quality controls, improved background handling options, auto-adapts to different models: Flux, SDXL and now gpt-image-1

Refactored

  • Moved subpackage plugin to the same level as cogt within pipelex for better visibility
  • Major cleanup in the unit tests, hierarchy significantly flattened
  • Strengthened error handling throughout inference flows and template preprocessing
  • Added make test-quiet (shorthand tq) to Makefile to run tests without capturing outputs (i.e. without pytest -s option)
  • Stopped using Fixtures for pipe_router and content_generator: we're now always getting the singleton from pipelex.hub

Fixed

  • Perplexity integration - Fixed breaking changes from recent updates

Dependencies

  • Added pytest-xdist to run unit tests in parallel on multiple CPUs. Not yet integrated into the Makefile, so run it manually with pytest -n auto (without inference) or pytest -n auto -m "inference" (inference only).
  • Swapped pytest-pretty for pytest-sugar - because readable test names > pretty tables
  • Updated instructor to v1.8.3
  • All dependencies tested against Python 3.10, 3.11, 3.12, and 3.13

Tests

  • TestTemplatePreprocessor
  • TestImggByOpenAIGpt
  • TestImageGeneration
  • TestPipeImgg

[v0.2.9] - 2025-05-30

  • Include pyproject.toml inside the project build.
  • Fix ImggEngineFactory: image generation (imgg) handle required format is platform/model_name
  • pipelex cli: Added list-pipes method that can list all the available pipes along with their descriptions.
  • Use a minimum version for uv instead of a fixed version
  • Implement AGENTS.md for Codex
  • Add tests for some of the tools.misc
  • pipelex cli: Rename pipelex run-setup to pipelex validate

[v0.2.8] - 2025-05-28

  • Replaced poetry by uv for dependency management.
  • Simplify llm provider config: All the API keys, urls, and regions now live in the .env.
  • Added logging level OFF, prevents any log from hitting the console

[v0.2.7] - 2025-05-26

  • Reboot repository

[v0.2.6] - 2025-05-26

  • Refactor: use ActivityManagerProtocol, rename BaseModelTypeVar

[v0.2.5] - 2025-05-25

  • Add custom LLM integration via OpenAI sdk with custom base_url

[v0.2.4] - 2025-05-25

  • Tidy tools
  • Tidy inference API plugins
  • Tidy WIP feature ActivityManager

[v0.2.2] - 2025-05-22

  • Simplify the use of native concepts
  • Include "page views" in the outputs of Ocr features

[v0.2.1] - 2025-05-22

  • Added OcrWorkerAbstract and MistralOcrWorker, along with PipeOcr for OCR processing of images and PDFs.
  • Introduced MissionManager for managing missions, cost reports, and activity tracking.
  • Added detection and handling for pipe stack overflow, configurable with pipe_stack_limit.
  • More possibilities for dependency injection and better class structure.
  • Misc updates including simplified PR template, LLM deck overrides, removal of unused config vars, and disabling of an LLM platform id.

[v0.2.0] - 2025-05-19

  • Added OCR, thanks to Mistral
  • Refactoring and cleanup

[v0.1.14] - 2025-05-13

  • Initial release 🎉