vibekanka

Turns transcribed Daggerheart session recordings into campaign canon on a self-hosted Kanka-CE instance, with Foundry VTT and GM-authored setting documents as additional ground-truth sources, and a Discord bot in the loop for anything too ambiguous to commit automatically.

Pipeline, per transcript: a session is recorded per-participant and transcribed upstream (see CLAUDE.md), so speaker labels are ground truth rather than a diarization guess, then ingested and classified statement-by-statement into canon facts, ambiguous facts, and out-of-character chatter. Canon facts are written straight to Kanka-CE. Ambiguous facts get posted as an @mention in a shared Discord channel for the relevant player or the GM to confirm before being committed.

Foundry VTT (actors, items, and journal entries) and a GM-provided setting document can also be synced in as ground truth alongside the transcript classifier — see CLAUDE.md for the full architecture and design decisions.

Setup

pip install -e ".[dev]"
cp .env.example .env  # fill in Kanka-CE, Discord, and LLM credentials

LLM calls go to the Anthropic API by default. To run them against a self-hosted Ollama instance instead, set LLM_PROVIDER=ollama plus OLLAMA_BASE_URL/OLLAMA_MODEL — see .env.example for the context-window caveat, which matters a lot for full-session transcripts.

Usage

Syncing sources into Kanka-CE:

vibekanka process <transcript>   # run one transcript through ingest -> classify -> sync
vibekanka foundry-sync           # sync Foundry VTT actors/items/journals into Kanka-CE
vibekanka sync-setting <doc>     # parse a GM-provided setting document into Kanka-CE entities
vibekanka sync-clarifications    # write clarified (human-answered) facts to Kanka-CE
vibekanka bot                    # long-running Discord bot for clarification replies

Reviewing and repairing what got synced (all conservative — verify/audit only report; the merge/rename/reset commands act only on what you name):

vibekanka verify                 # flag likely duplicate Kanka-CE entities from synced facts, for review
vibekanka audit                  # full-DB consistency check across all synced sources (add --llm for semantic review)
vibekanka merge-entities         # merge two Kanka-CE entities found by verify/audit
vibekanka rename-entity          # rename a Kanka-CE entity (e.g. a misheard proper noun)
vibekanka reset-session          # undo one session's Kanka-CE writes so it can be reprocessed
vibekanka backfill-subtypes      # fill in missing Kanka-CE type sub-categories on existing entities
vibekanka repair-entries         # rebuild entity entry text from the local facts table

Proper-noun glossary, used to bias both the classifier and upstream transcription:

vibekanka export-glossary        # draft a glossary from live Kanka-CE state, for hand-validation
vibekanka export-whisper-prompt  # trim a validated glossary towards WhisperX's 224-token --initial_prompt budget

Choosing a local model (see CLAUDE.md's "LLM providers"), when running classification against a self-hosted Ollama instead of the hosted API:

vibekanka eval-ollama <transcript>  # run model(s) N times over a real transcript and report
                                    # per-attempt failure rate, timing, token counts and
                                    # glossary compliance. Makes real LLM calls -- a sweep
                                    # over a local model can take hours.

Development

pytest          # run tests
ruff check .     # lint

Observability

Logging and Prometheus metrics for everything that leaves the process (LLM calls with estimated cost, Kanka-CE/Foundry HTTP calls) plus pipeline and Discord-bot progress counters. vibekanka bot is long-lived and exposes a normal pull endpoint on METRICS_PORT; one-shot commands instead push to a Prometheus Pushgateway (PUSHGATEWAY_URL) as they exit, since a short run usually finishes before a scrape could ever reach it. --metrics-port overrides both for any command. Both are unset by default.

docs/vibekanka-grafana-dashboard.json is a sample dashboard covering all of it. ansible/ and docs/*-runbook.md automate the self-hosted Kanka-CE, Foundry relay, and Pushgateway services this project talks to.

Status

ingestion/loader.py, analysis/classifier.py, kanka/client.py, storage/db.py, pipeline/orchestrator.py, pipeline/clarifications.py, pipeline/foundry_sync.py, pipeline/setting_sync.py, pipeline/audit.py, pipeline/glossary.py, and discord_bot/bot.py are implemented; the transcript pipeline, Foundry sync, setting-document sync, audit, and glossary export have all been live-tested end-to-end against real Kanka-CE (and, for Foundry, a real Foundry VTT instance) data. The Discord bot has likewise been live-tested against a real Discord server. The Ollama LLM backend is unit-tested only — there's no Ollama instance on this box to try it against. See CLAUDE.md for full architecture notes, known gaps, and design decisions.