Convo docs

Go from signing in to your first exported conversations using the defaults convo ships with.

• Generate your first conversations
• Review and export them

Every tab can walk you through itself. Press Tutorial in the top bar for a guided tour of the page you're on.

• Take a tour of a page you've never used
• Choose your path
• Stop any time

Save a run's recipe once (prompt set, models, counts, length mix and filters), then start runs from it.

• Make a cheaper copy of a pilot config
• Change the length mix
• Tidy up configs you no longer use

Start a run from a config with a cost estimate up front, watch it live, pause it, and read why conversations were rejected.

• Start a run and check the cost first
• Pause a run and pick it up later
• Find out why conversations were rejected

See exactly which personas, moods, occasions and lengths a run would plan, before you spend anything on model calls.

• Check a run's persona spread before spending money
• Plan seeds you can replay for a comparison

Connect the model servers convo writes and judges with, check they answer, and compare model prices before you pick one.

• Add an Ollama server on another machine
• Compare model prices before choosing one
• Connect a server that needs an API key

Write, copy, preview and roll back the pieces of text that make up the generation prompt.

• Make the AI friend's voice warmer
• Roll back a change with Versions and Diff
• Write a new occasion from scratch

Pick one segment for each slot of the prompt, or let each conversation draw its own, and preview the whole prompt.

• Build a prompt set from Baseline and swap two slots
• Let each conversation pick its own occasion
• Compare two register segments side by side

Keep the pool of user cards that conversations are seeded from accurate, varied and up to date.

• Fix a persona whose job is too vague
• Top up the pool with generated personas
• Take a persona out of rotation

Shape the AI friend's personality with voices, and retire one without touching past runs.

• Add a calmer, low-energy friend
• Retire a voice without breaking past runs

Edit the weighted lists that seeds draw from and the phrase lists the filters check, and validate before you save.

• Add an occasion and check its weight
• Catch a support-bot phrase the filter misses

Search every conversation from every run, fix text in bulk with a preview, and undo a whole batch at once.

• Find every conversation that overuses a word
• Replace a word in assistant turns, then undo it
• Pick up where another reviewer left off

Step through conversations one at a time with the keyboard, read them blind, and label them for the judge.

• Work through a run's flagged conversations
• Read blind, then label for calibration

Check how often the judge agrees with people, try stricter thresholds as a what-if, and re-cut a run from stored scores without regenerating.

• Check whether the judge agrees with people
• Try a stricter threshold and re-cut a run
• Test a filter phrase before adding it to a pool

Turn accepted conversations into JSONL training files, and build blind-read packets to check they sound like a real friend.

• Export accepted conversations for training
• Export the whole corpus with its persona and voice cards
• Compare two prompt sets with a blind read