Skip to content

Changelog

All notable changes to PsychScanner are documented here.

0.2.0 (2025)

New features

  • FeedbackBase API — subclass FeedbackBase and implement on_response(trial, response) to inject trial-level feedback into the conversation. Replaces the legacy generate_feedback interface. on_response receives the raw trial dict and a pre-parsed response dict; return a JSON string to inject or None to skip.

  • Session checkpointing — SessionTunnel writes BEGIN, SCAN, and END log entries to a JSON file alongside experiment data. Set tunnel_status="1" on ExpCardInit to enable. Re-running the same script after an interruption resumes from the last completed participant.

  • save_expcard / load_expcard — serialize and reload a complete experiment configuration, including inline task JSON and persona JSON, to a portable file. Designed for reproducibility and sharing across machines.

  • get_task_template() — returns an empty task JSON skeleton for quick scaffolding.

  • Mock LLM provider — use model="mock-chat-model" and family="mock-llm" for zero-API-key testing of pipelines.

  • parser_raw flag — set parser_raw=True to store the raw AIMessage alongside the parsed response dict in trial["pred_resp"]["_raw"].

  • summary_k context compression — when context overflows in Convo mode, automatically summarize the oldest summary_k messages into a single summary message.

  • Per-trial parser dispatch — pass a callable to parser that takes a trcode string and returns the appropriate parser class. Enables multi-phase tasks with different output schemas per phase.

Provider support additions

  • Google / Gemini (google, gemini, google-genai family aliases)
  • Together.ai (together)
  • Fireworks.ai (fireworks)
  • Cohere (cohere)
  • Azure OpenAI (azure)

Documentation

  • New MkDocs Material documentation site at https://saurabhr.github.io/psychscanner/
  • Full configuration reference, API reference, and user guides
  • Example notebooks updated to use current API

Breaking changes

  • FeedbackBase.generate_feedback(trdata, pred_dict, input_dict, parser_status, trial_item_collector) is removed. Replace with on_response(trial, response).
  • memory_type is not a valid field; use memory with values "SingleTurn" or "Convo".
  • system_prompt, task_prompt, and response_format are not fields on ExpCardInit. Task content is set via task_file; personas via persona_files.

0.1.0 (2024)

Initial release

  • ExpCardInit / ExpCard — experiment configuration via Pydantic model
  • ScannerModel — runs experiment trials against LangChain chat models
  • SingleTurn and Convo memory modes
  • Bundled parsers: DefaultLiteralAgree, DefaultLiteralVivid15, DefaultLiteralVivid010, DefaultResponseRating, Response_part_1_rm, Response_part_2_rm
  • Built-in VVIQ-16 imagery questionnaire as the default task
  • to_csv / concat_csv export utilities
  • Support for OpenAI, Anthropic, Groq, Mistral, HuggingFace, and Ollama providers