ChangeLog

All notable changes to SDialog will be documented here.


[Unreleased]

Added

  • sdialog.server:

    • agent.serve(chat_ui=True) also serves a web chat GUI at /chat, with several conversations saved in the user’s browser, light and dark themes, and a compact layout when embedded in other pages, with an <iframe> or with the floating chat button of /chat/widget.js. ChatUI(...) sets its path, title, single or multiple conversations, Markdown replies, whether to show the agent’s reasoning and tool calls, and which sites can embed it.

Changed

  • sdialog.evaluation:

    • ReadabilityScore and ConversationalFeatures name their results with dashes, like the feature names used to request them (e.g. "flesch-reading-ease", "question-rate" instead of "flesch_reading_ease", "question_rate"), and their default names are "readability-score" and "conversational-features".


[0.8.3] 2026-11-08 🎧

Added

  • sdialog.audio:

    • ButReverbDB(path, download=True) downloads BUT ReverbDB (8.7 GB) and extracts it into path the first time (d917170).

    • audio_format="flac" in to_audio() and dialogs_to_audio(), to save lossless FLAC files (about half the size of WAV) instead of WAV (d917170).

  • Tutorials:

    • The audio tutorials keep their audio players, so they can be listened to without running them (c6ef098).

Changed

  • sdialog.audio:

    • DialogAudio.display() embeds the audio in the notebook as MP3, about 7 times smaller than WAV, so notebooks saved with their outputs stay small. display(audio_format="wav") embeds the exact samples, as before (d917170).


[0.8.2] 2026-11-05 🎙️

Added

  • sdialog.audio:

    • ButReverbDB, to reverberate dialogues with real rooms measured in BUT ReverbDB (downloaded separately), e.g. dialog.to_audio(reverb=ButReverbDB(path), noise="room"). With noise="room", the background noise is the one recorded by the same microphone in the same room. See Measured Rooms (9a2b888).

Fixed

  • sdialog.audio:

    • dialogs_to_audio() on a single GPU ran out of memory after dialog.to_audio() (e.g. in the audio tutorials): it loaded a second Dia2 model in the same process. It now reuses the one already loaded (b8545e1).


[0.8.1] 2026-11-04 🩹

Fixed

  • Dependencies:

    • Local Hugging Face models (e.g. for interpretability) failed to load unless accelerate was installed. It is now installed with SDialog (cd23cb3).

  • sdialog.agents:

    • An agent without persona or system prompt that speaks first (e.g. in talk_with()) now raises a clear ValueError when its model can’t generate from an empty input (e.g. OpenAI, Ollama), instead of an unclear error or an empty answer (cab5e44).

  • LLMs:

    • Ollama gpt-oss models returned empty answers for structured outputs (e.g. DialogGenerator, PersonaGenerator, LLM judges), and agents using them failed to start. Both now work (64117b2).

  • sdialog.evaluation:

    • dialog2graph() failed to label nodes when the model name had a provider prefix (e.g. "ollama:qwen3:14b", "openai:gpt-4.1"). Any model supported by SDialog now works; labels may differ slightly from previous versions (cd23cb3).

  • sdialog.audio:

    • HuggingfaceVoiceDatabase, the default voices of dialog.to_audio(), took about 18 GB of RAM to load. It now decodes each voice only when used and loads in about 2 seconds; it also supports datasets>=4 (5e16407, f52c710).

    • dialogs_to_audio() overwrote the files of dialogues with the same id (e.g. [dialog] * 3). They are now saved as <dialog_id>_<position> (4121dea).

    • When Dia2 fails capturing CUDA graphs (e.g. on H100 GPUs), Dia2TTS now suggests Dia2TTS(use_cuda_graph=False) instead of raising cryptic cuDNN errors (e3817bf).


[0.8.0] 2026-11-03 🏠

Added

  • sdialog.whatsapp:

    • Users now see “typing…” while the agent writes its reply, refreshed every 20 seconds for long replies (Meta hides it after 25). Disable it with WhatsAppConfig(typing_indicator=False); like the read receipt it comes with, it is only shown with mark_as_read=True (the default).

    • on_event argument of serve_whatsapp(): a function called as on_event(user_id, event) with each event of a user’s turn (orchestrator instructions, tool calls and the reply) once the reply is sent, e.g. to alert a person when a safety orchestrator fires.

  • sdialog.audio:

    • Room, to choose the room simulated by dialog.to_audio(reverb=Room(...)), as the legacy pipeline’s Room allowed but reworked to run without the legacy pipeline’s dependencies: it only needs pyroomacoustics, already installed with sdialog[audio] (no dSCAPER or system packages). It sets the dimensions, RT60 or surface materials, microphone position (coordinates, ceiling, wall, between the speakers or on a speaker’s chest), microphone directivity (cardioid, figure-eight, …) and where it points, and speaker positions (random when not given). Rooms can also be given per speaker (e.g. reverb={"Agent": Room(...), "Caller": Room(...)}), and room.compute_rirs() returns the impulse responses. See Custom Rooms.

    • Documentation: Coming from the Legacy Audio Pipeline, with the old API’s equivalents in dialog.to_audio() and how to install the last versions that include it (sdialog<0.7).


[0.7.0] 2026-10-30 🧪

Added

  • sdialog.evaluation:

    • forbidden_tools argument of ToolSequenceValidator: tools that must not be called anywhere in the dialogue, regardless of the order (e.g. ToolSequenceValidator(["get_service_plans"], forbidden_tools=["verify_account"])).

Fixed

  • sdialog.audio:

    • Qwen3TTS and Qwen3TTSVoiceClone no longer modify the tts_pipeline_kwargs dictionary they are given (e.g. PerTurnTTS‘s tts_kwargs), nor their shared default one: the language, sampling and reference audio of one call were left in it.

    • DocumentFormatNormalizer removed newlines and tabs instead of replacing them with spaces, gluing words together (e.g. “Hello\nworld” became “Helloworld”).

  • sdialog.generators:

    • Paraphraser.prompt() returned "<Template memory:...>" instead of the instruction template.

    • PersonaGenerator with a persona class whose fields are required (no default): int, float and bool fields were never generated (nor set by generated_attributes rules such as "{{20-25}}"), keeping their 0 or False placeholder.

  • sdialog.interpretability:

    • Steering added to an inspector after attaching it to the agent (e.g. agent = agent | inspector, then inspector + direction or steerer + inspector, as in the DirectionSteerer example) was ignored when the inspector had no steering function yet. Adding it before attaching (agent | inspector + direction, as in the tutorials) already worked.

    • Inspector(target, agent=agent) raised an AttributeError; it is now attached like agent | inspector.

    • Adding an inspector to an agent without a local Hugging Face model now raises a clear RuntimeError (it raised an AttributeError).

  • sdialog.orchestrators:

    • SimpleResponseOrchestrator with top_k=1: the suggested response was split into characters (“(1) W; (2) h; …”), or failed with an AttributeError when responses are given as a dictionary of actions.

  • sdialog.evaluation:

    • ConversationalFeatures (and the scores based on it) and ReadabilityScore (and the scores based on it) now ignore the speaker names: they were computed on the “Speaker: text” lines, so BackChannelRateScore was always 0 and speaker names were counted as words (e.g. in LexicalDiversityScore or GunningFogScore).

    • Back-channel turns followed by punctuation (e.g. “Okay.” or “Yeah!”) are now detected by BackChannelRateScore.

    • DialogFlowPPL and DialogFlowScore (and dialog2graph()) ignored ai_speaker (system_speaker_name): the speaker of the first turn was always taken as the system, so the speakers were swapped in dialogues started by the user. With use_only_ai_speaker=True, the graph was then built from the user’s turns and scoring failed with a KeyError.

    • StatsEvaluator.plot() failed with matplotlib >= 3.11 (boxplot() no longer accepts labels).

    • Dataset evaluators given the path of a dialogue file or folder (e.g. MeanEvaluator(score)("dialogs/")) iterated over the characters of the path instead of loading the dialogues.

    • DatasetComparator with datasets given as a list (e.g. comparator([dialogs_a, dialogs_b])) crashed with a TypeError in score and embedding evaluators.

    • ToolSequenceValidator with "not:" tools: the check was inverted (e.g. with ["not:send_receipt", "charge_payment"], sending the receipt after the charge failed and sending it before passed). A "not:" tool now must not be called at its position in the sequence, between the previous and the next required tools: ["not:x", "a"] means no x before a, ["a", "not:x"] no x after a, and ["not:x", "a", "not:x"] no x at all.

    • LLMJudgePersonaAttributes could not be created (AttributeError: its prompt template was rendered as a string). The persona and speaker are now passed to the prompt when judging.

    • LLMJudgeScore.judge() with several dialogues: the prompt asks for one score per dialogue, but the output schema only accepted a single score, so models with structured output returned one score for all the dialogues (and a list failed validation). Lists of scores (and reasons) are now accepted, each score within [min_score, max_score].

  • sdialog.whatsapp:

    • Messages that reach the server long after they were sent are now ignored instead of answered out of context. Meta retries webhooks it could not deliver (e.g. while the server or tunnel was down) for days, so users could get replies to hours-old messages. The threshold is max_message_age in serve_whatsapp() (120 seconds by default; None to answer all messages).

Removed

  • sdialog.audio:

    • The legacy audio pipeline (sdialog.audio.pipeline.to_audio / AudioPipeline, dSCAPER-based), deprecated since 0.5.0, with the modules only it used: dialog (AudioDialog), turn, room, room_generator, acoustics_simulator, room_acoustics_backends, impulse_response_database, processing, dscaper_utils, jsalt, legacy and utils. dialog.to_audio() now raises a TypeError for the legacy pipeline’s arguments (e.g. perform_room_acoustics=True) instead of routing them to it; use its own options (preset=, reverb=, noise=, channel=, …) instead.

    • The sdialog[audio-legacy] extra, and the audio tutorials 1-10 that used the legacy pipeline (tutorials/01_audio/0.quickstart.ipynb covers dialog.to_audio()).


[0.6.0] 2026-10-29 💬

Added

  • sdialog.whatsapp (#135, contributed by @amitbhakt):

    • Agent.serve_whatsapp() (and Server.serve_whatsapp()) to let real users chat with agents over WhatsApp through the Meta WhatsApp Cloud API. Each user (phone number) has their own conversation, saved in SQLite and restored on their next message. See Serving Agents on WhatsApp.

    • Webhook requests are verified with the app secret (required), duplicate deliveries from Meta are ignored, and non-text messages get a short “not supported” notice.

    • Per-user rate limits: max_messages_per_minute (6 by default) and max_messages_per_hour (60 by default); extra messages are dropped without calling the LLM, with a single notice.

    • Up to max_workers users (8 by default) are served at the same time, each on a clone of the agent; a user’s messages are always processed one at a time, in order.

    • Conversations start over when the user sends /reset, after session_timeout seconds of inactivity, or, if welcome messages are enabled on the phone number, when a user opens a chat for the first time (Meta’s request_welcome event; the agent greets them). With the default session_timeout="auto", the timeout is 30 minutes, or 24 hours if welcome messages are enabled (detected at startup); None disables it.

    • session_retention_days (30 by default): conversations of users inactive for longer are deleted from the database.

  • sdialog.agents:

    • Agent.get_state() / Agent.set_state() to save a conversation (memory, finished flag and orchestrator state) and restore it later, e.g. on a fresh or reset agent.

    • Agent.clone() to copy an agent with its own conversation (memory, orchestrators, LLM settings such as the seed) that shares its language model, so weights and API clients are not duplicated.

  • sdialog.agent_pool:

    • AgentPool to run up to max_workers conversation turns of an agent at a time (8 by default), each on a fresh clone, with turns of the same conversation kept in order. Agents with a local Hugging Face model or inspectors run one turn at a time. Used by both servers.

  • sdialog.server:

    • max_workers in Agent.serve() and Server.serve() / serve_async() / serve_in_thread() to process several requests at the same time (see Changed).

    • Per-client (IP address) rate limits of chat requests: max_requests_per_minute (30 by default) and max_requests_per_hour (300 by default). Extra requests get HTTP 429 with a Retry-After header; pass None to disable a limit (e.g. for local scripts).

  • sdialog.orchestrators:

    • get_state() / set_state() hooks for stateful orchestrators; ChangeMindOrchestrator saves its trigger counter.

  • sdialog.rate_limit:

    • RateLimiter, a per-key sliding window rate limiter with several limits at once, used by both servers.

Changed

  • sdialog.server:

    • Served agents now process up to 8 requests at the same time (stateless mode, the default) instead of one, each on a clone of the agent. The agent’s tools and custom functions (e.g. postprocess_fn) may therefore run concurrently and must be thread-safe; pass max_workers=1 for the previous behavior. Agents with a local Hugging Face model or inspectors, and stateful agents, still process one request at a time.

    • Chat requests are now rate limited by default (30/minute and 300/hour per client IP). Scripts that call a served agent in a fast loop (e.g. evaluations) may get HTTP 429 errors; pass max_requests_per_minute=None, max_requests_per_hour=None to serve() to disable the limits.

Deprecated

  • sdialog.audio:

    • The legacy audio pipeline (sdialog.audio.pipeline.to_audio / AudioPipeline, dSCAPER-based), deprecated since 0.5.0, and the sdialog[audio-legacy] extra will be removed in an upcoming release, together with the tutorials that use them. Use dialog.to_audio() instead.

Fixed

  • sdialog.server:

    • The REST API server no longer blocks while an agent generates a reply: generation runs in a worker thread, so other requests (e.g. /health, /v1/models) are answered meanwhile instead of waiting for it.


[0.5.4] 2026-10-26 🩹

Fixed

  • sdialog.audio:

    • Dia2TTS no longer crashes with a CUDA “index out of bounds” error on long generation windows, which also left the GPU process unusable (every later dialogue failed too). The reference voices took up part of Dia2’s context, so long windows ran past its position limit. Generation now stops at that limit; a window cut short is regenerated one turn at a time.

    • The rotary position (RoPE) table of Dia2 is now sized to the positions generation can actually reach (prefix_len + max_context_steps) instead of max_context_steps only (#136). With reference voices, generation runs at absolute positions shifted by the prefix length, so windows were still cut short (or, before the stop above, crashed) even when they fit in the generation budget. The table grows only when needed and existing rows are unchanged, so generations that already fit give the same output. The stop above stays as a safety check against the new table size.

    • Reference voices given as audio files (e.g. from a local voice database) are now cut to reference_seconds (8 s by default), as arrays already were. Long reference files took up most of Dia2’s context, leaving little room for the generated audio and making the crash above frequent.


[0.5.0] 2026-10-15 🔊

Added

  • sdialog.audio:

    • New dialog.to_audio() engine (sdialog.audio.synthesis): in-memory dialogue-to-audio synthesis returning a DialogAudio object with the waveform and ground-truth annotations (turn segments, word timestamps, RTTM/JSON export, save()).

    • Dia2 dialogue TTS (Dia2TTS) as the default engine, vendored under sdialog.audio.tts.dia2 (Apache-2.0). Changes from upstream: seeding, Whisper loaded once, and in-memory reference audio.

    • Voice selection with voices="db" | "design" | "random", or an explicit {speaker: voice} dictionary.

    • Realism controls: pauses, overlap, reverb (simulated rooms or measured RIRs), noise, channel (telephone codecs) and level_jitter_db, either shared or per speaker. Output in mono or stereo (one channel per speaker).

    • Scene presets: clean, conversational, room, call_center, far_field, hard, plus register_preset() for custom ones.

    • dialogs_to_audio() for parallel, multi-GPU generation, with one model per worker.

    • BaseDialogTTS interface for dialogue-level engines, and PerTurnTTS to use classic engines (Kokoro, Qwen3-TTS, …) with dialog.to_audio().

    • New sdialog[audio-legacy] extra with the dependencies of the legacy pipeline.

Changed

  • sdialog.audio:

    • sdialog[audio] now installs the lightweight new engine plus Dia2’s dependencies (Python >= 3.10). dSCAPER, librosa, sox and the other legacy dependencies moved to sdialog[audio-legacy].

    • import sdialog.audio no longer imports heavy dependencies; legacy names are loaded lazily.

    • The legacy helpers that lived in sdialog/audio/__init__.py moved to sdialog.audio.legacy. They can still be imported from sdialog.audio.

Deprecated

  • sdialog.audio:

    • The legacy pipeline (sdialog.audio.pipeline.to_audio / AudioPipeline, dSCAPER-based) is deprecated. dialog.to_audio() still routes calls that use its arguments (e.g. perform_room_acoustics=True) to it, with a DeprecationWarning.

Fixed

  • sdialog.audio:

    • Dialog.to_audio() no longer hides real import errors behind “module not installed”, and the documented perform_tts argument no longer raises a TypeError.

    • Persona ages given as text (e.g. “middle-aged”) no longer break voice selection.

    • audio/prompts/*.txt is now included in the package.

    • Removed leftover debug prints in the audio module.


[0.4.7] 2026-04-14 📞

Added

  • sdialog.audio:

    • New telecommunication simulation engine and codecs support.

    • New modular room acoustics backend system (pyroomacoustics and telecommunications backends).

    • New text/audio normalization utilities in sdialog.audio.normalizers.

    • New prompt templates for sound-effect annotation with and without stage-aware context.

    • New tutorials for telecommunication simulation, overlaps/pauses, and sound events.

Changed

  • sdialog.audio:

    • Refactored acoustics simulation to a modular backend architecture.

    • Extended audio pipeline with overlap/pause control, sound-event handling, and improved post-processing.

    • Improved normalization behavior for Qwen3-TTS-based generation and voice cloning.

Fixed

  • sdialog.generators:

    • Improved validation error handling in Paraphraser.

  • sdialog:

    • Prevented API keys from being persisted in LLM metadata.

  • sdialog.audio:

    • Fixed sentence-transformers modality mismatch issues in audio workflows.

    • Fixed multiple robustness issues in audio processing (NumPy conversion, dScaper edge cases, dataset handling, and normalization consistency).

    • Fixed audio test/tooling compatibility issues (torchcodec/qwen_tts stubs and type hints).


[0.4.6] 2026-03-12 🧠

Added

  • sdialog.agents:

    • New @final_response_tool decorator to return tool output directly as the final agent response (bypassing post-tool LLM synthesis for non-empty outputs).

    • Example:

      from sdialog.agents import Agent, final_response_tool
      
      @final_response_tool
      def get_table(callsign: str) -> str:
          return "| Timestamp | Transcription |\n|---|---|\n| ... | ... |"
      
      agent = Agent(tools=[get_table])
      

Changed

  • sdialog.evaluation:

    • KDEDistanceEvaluator now supports multiple values per dialog.

    • Improved placeholder handling in STAR response generation.

    • Implemented Laplace smoothing in dialog log-likelihood computation.


[0.4.5] 2026-02-18 🎙️

Added

  • sdialog.audio:

    • Voice cloning support in voice databases.

    • New TTS classes: Qwen3TTS, Qwen3TTSVoiceClone, and BaseVoiceCloneTTS.

    • Persona-based dynamic voice generation and customizable voice descriptions.

    • Improved speaker name / role mapping for voice assignment.

Changed

  • sdialog.audio:

    • Dialog.to_audio() supports output path as positional argument.

    • Audio pipeline now supports more flexible voice assignment.

    • Qwen3-TTS voice cloning set as default dynamic voice creation path.

Fixed

  • sdialog.agents:

    • Restored orchestration behavior when agent is served.

  • sdialog.audio:

    • Fixed speaker-role mapping and case-sensitivity edge cases.

    • Fixed optional audio path handling in Dialog.to_audio().

    • Fixed label assignment and default keep_duplicate handling.

  • sdialog.evaluation:

    • Fixed division-by-zero / log(0) edge cases in divergence computations.

    • Added jitter handling for KL/CS divergence when variance is zero.

    • Fixed fallback behavior for unknown edges in flow score computation.


[0.4.3] 2025-11-30 📊

Added

  • sdialog.evaluation:

    • New readability and conversational metrics.

    • New scoring classes for individual linguistic features.

    • Added Dialog.from_folder() to load dialogs from directories.

    • Added ToolSequenceValidator for validating tool usage order in dialogs.

Changed

  • sdialog.evaluation:

    • Enhanced plotting support with customizable titles and labels.

  • sdialog.interpretability:

    • Improved top-k visualization and steering defaults.

Fixed

  • sdialog.agents:

    • Improved tool invocation error handling.

  • sdialog:

    • Updated dialog prompt formatting.


[0.4.2] 2025-11-21 🛠️

Changed

  • sdialog.agents:

    • Improved progress bar behavior during dialog generation.

  • sdialog.audio:

    • Updated dependencies and dscaper integration behavior.

Documentation

  • Added clearer Kokoro TTS installation instructions.


[0.4.1] 2025-11-20 🔧

Fixed

  • sdialog.audio:

    • Fixed dscaper timeline generation bug and test updates around deleted directories.

Documentation

  • Updated audio tutorials and README installation instructions (including sox/ffmpeg and setup guidance).

[0.4.0] 2025-10-30 🔊

Added

  • sdialog.audio: New audio generation module to transform text dialogs into audio conversations

    • Dialog.to_audio(): Convert any text dialog to audio with a single method call

    • Text-to-Speech (TTS): Support for multiple TTS engines including Kokoro and Hugging Face models

    • Voice databases: Automatic or manual voice assignment from Hugging Face, local storage, or on-the-fly generation based on persona attributes (age, gender, language)

    • Acoustic simulation: Realistic room acoustics simulation with ray tracing technology for spatial audio

    • Microphone simulation: Professional microphone impulse responses from brands like Shure, Sennheiser, and Sony

    • Room generation: Define and generate diverse room types with customizable properties (dimensions, wall materials, furniture placement)

    • Multiple formats: Export to WAV, MP3, or FLAC with custom sampling rates

    • Background/foreground effects: Add environmental sounds and acoustic variations

    • 7 comprehensive tutorials covering audio generation, acoustic simulation, rooms, voice databases, and impulse responses

  • sdialog.interpretability:

    • Support for inspecting layer and component inputs with Inspector(target="model.layers.15", inspect_input=True)

    • Support for inspecting and steering input tokens (tokens given as input, not only generated ones), e.g., inspector.input[i][j].act where i is turn index and j is input token index

  • sdialog.config:

    • Support for Anthropic backend (#100)

    • Support for Azure OpenAI backend (#100)

Fixed

  • sdialog.agents: Agent memory reset when no prompt system and no persona is given


[0.3.3] 2025-10-30 🚀

Added

  • sdialog.server:

    • New module to serve agents via an Ollama/OpenAI-compatible REST API (works with UIs like Open WebUI) (#92)

  • sdialog:

    • Dialog.from_huggingface() to load/download dialogues directly from Hugging Face datasets (#59)

Changed

  • sdialog.evaluation:

    • LLM judge methods now accept additional user-defined template arguments (e.g., like document in this example) (#86)

  • sdialog.agents:

    • Improved Agent initialization so agents can act as a proxy for external conversational systems (#90, fa1d8f3)

Fixed

  • sdialog.evaluation:

    • Corrected Flesch Reading Ease and Gunning Fog score calculations (d1d4260)


[0.3.0] 2025-10-03 ✨

Added

  • sdialog:

    • Context: new class class to explicitly model the common/shared context of conversations (#73)

    • Dialog: merge functionality - Added option to merge consecutive turns of the same speaker when loading a dialog (#77)

    • Dialog: built-in string support - Added support to built-in str functions for Dialog class (#83)

  • sdialog.agents: Added new sdialog.agents module and moved Agent class inside (#81)

    • Agent: thinking capabilities - Agents can now handle internal thinking processes (#95)

    • Agent: tools support - Added tools capabilities to Agents (e.g. RAG or any other function) (#84)

      • New tutorial for agents with tools and thoughts.

  • sdialog.generators:

    • ContextGenerator: new class added to explicitly model the common/shared context of conversations (#73)

    • Paraphraser: new class class to paraphrase dialogues (#76)

  • sdialog.evaluation:

    • LinguisticFeatureScore: new class added to compute Flesch reading ease, Gunning fog, Hesitation rate, and/or Mean turn length (#63)

  • sdialog.personas:

    • Customer and SupportAgent: new personas added for customer service dialogues (#85)

    • Persona: Added static method to get the list of all attributes in Persona class (#79)

Changed

  • sdialog: Improved metadata handling (#66)

  • sdialog.interpretability: Improved and simplified the way inspection targets are defined in interpretability submodule (#78)

  • sdialog.evaluation.base:

    • LLMJudgeYesNoOutput: Renamed attribute yes to positive (#86)

    • LLMJudgeScoreOutput: Renamed attribute feedback to reason (#86)

Fixed

  • sdialog.generators: Fixed potential bug in PersonaDialogGenerator class (#67)

Enhanced

  • sdialog.agents: Added base_model attribute to Agent to direclty access the LLM’s underlying model for mechanistic interpretability (#74)

  • sdialog.config: Added clear_cache() method to config (#75)

Documentation

  • API Documentation: Refactored/cleaned all components and added docstrings with examples (#82, #88)

  • Updated all tutorials to work with new code and added “Open in Colab” badges

  • Completed API documentation for initial official release (#87)

  • Automatic generation of llm.txt from API documentation (24f6ee6)


[0.1.0] 2025-08-05 🌱

Added

  • Multi-backend support (Hugging Face, Ollama, OpenAI, AWS)

  • Enhanced persona generation (beyond initial PersonaDialogGenerator)

  • Interpretability module (sdialog.interpretability): inspectors, steerers, hooks, intruders

  • Evaluation module (sdialog.evaluation): metrics, LLM-as-a-judge scoring, evaluators, dataset comparators

Changed

  • Standardized / improved dialog format

Notes

  • 500 commits since 0.0.2 (post-JSALT 2025 consolidation)

Pending

  • Audio module (sdialog.audio) integration

  • Documentation updates


[0.0.2] 2025-06-03 🔧

Added

  • language attribute to Persona class

  • PersonaDialogGenerator to generators module to support persona-based dialogue generatin with single LLM:

    from sdialog.generators import PersonaDialogGenerator
    
    dialog_generator = PersonaDialogGenerator(
        model=MODEL_NAME,
        persona_a=bob_persona,
        persona_b=alice_persona,
    )
    
    dialog_generator.generate().print()
    

Fixed

  • Python 2 and 3 compatibility problem with scikit-learn (using version 0.20.1 from now on)

  • PyPi: setup.py: long_description_content_type set to 'text/markdown'


[0.0.1] 2025-05-22 🎉

(initial release)