1.4.0

What's Changed

• Add get_experiment_output tool for retrieving raw training logs by @nLamprok in #11

Full Changelog: 1.3.0...1.4.0

1.3.0

Key Changes

1. Query Functionality

• query_experiments(name, comparison, value, workspace, project_name)
  • Finds experiments that have a logged item matching specified criteria
  • Supports multiple comparison operators: ==, !=, >, >=, <, <=, contains, endswith, startswith
  • Supports type-specific queries using TYPE:name format where TYPE can be:
    • "Environment"
    • "Metadata"
    • "Metric"
    • "Other"
    • "Parameter"
  • Special value formats supported: true, false, none, datetime:ISO-FORMAT-DATETIME
  • Automatically handles numeric conversion and type disambiguation

2. Query Discovery Tools

• get_project_logged_item_names(workspace, project_name)

  • Returns a list of all named items that have been logged in a project
  • Helps users discover what fields are available for querying
  • Useful for understanding what data is available before constructing queries

• get_project_tags(workspace, project_name)

  • Returns a list of all tag texts that exist in a project
  • Helps users discover available tags for filtering experiments

3. Tag-Based Experiment Discovery

• find_tagged_experiments(tag, workspace, project_name)
  • Finds all experiments tagged with a specific text
  • Returns a list of experiment IDs matching the tag
  • Provides helpful error messages when no experiments are found

4. Enhanced Experiment Details

• get_experiment_details(experiment_id) - Enhanced
  • Now includes others field in the response
  • Returns a list of dictionaries with "other" names and values
  • Previously only returned metrics and parameters

5. CSV Export Functionality

• experiment_spreadsheet(workspace, project_name, experiment_keys)
  • Exports experiment details to CSV format
  • CSV files are made available as MCP resources (accessible via resource URI)
  • Exports include:
    • Workspace name
    • Project name
    • Experiment name
    • Duration (in seconds)
    • Metrics (as JSON string)
    • Parameters (as JSON string)
  • Can export all experiments from a project or specific experiments by ID
  • Generates timestamped filenames
  • Returns resource URI for accessing the CSV file without processing through LLM

6. Additional Improvements

• OpenTelemetry/Opik Logging

  • Added OpenTelemetry file and Opik logging support
  • Fixes for OpenTelemetry issues
  • Enhanced telemetry tracking for tool usage

• Sorting Fixes

  • Fixed sorting issues in experiment listing

Types of Questions That Can Now Be Answered

Based on the PR description, the following types of questions can now be answered:

1. Tag-based queries:

   • "Are there any experiments that are tagged in the 'sequential' project?"
   • "What experiments are tagged with 'marked' in this project?"

2. Value-based queries:

   • "Which experiment has a 'storage_size_bytes' of less than 66,386 in the 'sequential' project?"
   • "Are there any experiments that have a storage_size_bytes that is less than 66,000 bytes?"

3. Detailed experiment information:

   • "Can you tell me about 'identical_caviar_1647' experiment in project 'sequential'?"
   • Now includes "others" information that was previously missing

4. Data export and analysis:

   • Export experiment data to CSV for external analysis
   • Query experiments by any logged metric, parameter, or other value
   • Discover available queryable fields and tags in a project

Technical Details

Query Implementation

• Uses Comet API's get_query_variables() to discover available query fields
• Supports automatic type disambiguation when field names are unique
• Requires explicit type specification (TYPE:name) when field names exist in multiple types
• Handles value conversion (strings to numbers, boolean to 0/1, datetime parsing)

CSV Export Implementation

• Creates CSV files in memory using io.StringIO()
• Registers files as MCP resources via resource_manager.create_file()
• Returns resource URI for client access
• Handles errors gracefully, skipping experiments that fail to load

Error Handling

• All new functions include proper error handling
• Returns helpful error messages when:
  • Projects don't exist
  • No experiments match queries
  • Field names are ambiguous (requires type specification)
  • Experiments fail to load during export

Full Changelog: 1.2.0...1.3.0

1.2.0

What's Changed

• Add ca-certificates to Dockerfile for SSL/TLS certificate verification by @nLamprok in #4

Full Changelog: 1.1.0...1.2.0

1.1.0

What's Changed

• Update server docstring by @nLamprok in #2
• Add Docker Support for Comet MCP Server by @nLamprok in #1
• Add instructions to integrate MCP via docker by @nLamprok in #3

New Contributors

• @nLamprok made their first contribution in #2

Full Changelog: 1.0.4...1.1.0

1.0.4

Full Changelog: 1.0.3...1.0.4

1.0.3

Full Changelog: https://github.com/comet-ml/comet-mcp/commits/1.0.3