Studio
Open the demo: eight weeks of synthetic feedback in eleven languages from app stores, a help desk, an NPS survey, Bluesky and GitHub, with a login spike after an app update. No real people or customers are in it; the contact details and ID numbers in a few records are made up (checksum-valid, so redaction catches them).
Studio shows k-anonymous aggregates, stable themes with seven-day trends and a weekly sparkline
(themes whose last seven days at least doubled the seven before are marked Rising), a knowledge
graph that links themes to sources, languages, intents and sentiment, and redacted evidence for
each theme, with placeholders such as <EMAIL> highlighted. It never shows authors and never edits
data.
Privacy panel
Section titled “Privacy panel”Every Studio (live, export and demo) has a Privacy panel, built from aggregates only, so the viewer role sees it too:
- how often each redaction placeholder (
<EMAIL>,<PHONE>,<CARD>,<IBAN>,<IN_AADHAAR>,<BR_CPF>, …) occurs in stored text, and in how many records; - the k-anonymity threshold, and how many themes and source, language or label groups it hides;
- how many distinct pseudonymised authors there are (never who they are);
- the egress mode (
air_gapped,privateorhybrid) of the server, in live mode only.
Decisions
Section titled “Decisions”When classify stores fields (for example from a decision model
with team, urgency or yes/no fields), Studio shows a Decisions panel: k-anonymous counts per
answer (score levels in order) and the share of records marked for review because an answer was
below its confidence cutoff. Evidence lists each record’s labels with the model’s confidence.
Pipelines with routes also get a Routes chart: records per route rule.
How the demo is built
Section titled “How the demo is built”obsei demo builds the static demo from the synthetic records in obsei/demo.py: it redacts them
as ingest would, labels them, groups themes, and exports a snapshot marked demo. Only the demo
shows a short guided intro, before-and-after redaction cards (the raw text exists only because the
data is synthetic; Studio never has raw text otherwise) and a few Ask your data answers recorded
from the dataset, citing record ids shown in its evidence. Demo records are routed with the routes
of examples/decision-routing.yaml without its refunds rule (billing chosen by the demo’s team
field), using the same
routing code as a pipeline. In your deployment Ask uses your own
model, and agents get the same answers over MCP.
The demo’s labels (sentiment, intent, team, urgency and an angry yes/no) come from the default local
classifier, MoritzLaurer/bge-m3-zeroshot-v2.0, run once on a CPU and committed as
obsei/demo_labels.json, so the build needs no model. They are shown as the model produced them,
mistakes included: 21 of 136 records are marked for review, almost no message is judged angry and
most urgencies are “this week”, so the demo’s urgent-bugs route catches only 4 records, which
Studio hides because they are fewer than k (see
known limits). To label it again with the local classifier, or
with a decision model whose endpoint URL is in an environment variable:
uv run --extra local obsei demo --local-labels \ --save-labels packages/obsei/src/obsei/demo_labels.json --out docs/public/demoexport OBSEI_DEMO_DECISION_URL=... # the decision endpoint URL your server documentsuv run obsei demo --decision-url-env OBSEI_DEMO_DECISION_URL \ --save-labels packages/obsei/src/obsei/demo_labels.json --out docs/public/demoThe demo groups the same issue across languages with the multilingual model (needs network to Hugging Face once):
uv run --extra embeddings obsei models downloaduv run --extra embeddings obsei demo --embedder local --out docs/public/demoobsei demo follows OBSEI_EGRESS_MODE (air-gapped by default), so it loads the model only from
the local cache (OBSEI_MODELS_DIR, or the fastembed cache) that obsei models download fills.
--embedder takes the same names as themes.embedder: hashing, auto, local or local:<model>;
with the offline hashing embedder each language forms its own theme. The embedder that built the
demo is recorded as embedder in data.json. Demo themes carry curated labels; real themes are
labelled from shared keywords or by your labeler model.
obsei serve hosts Studio at /studio/. Studio asks for OBSEI_API_TOKEN and keeps it for the
browser tab only. The Ask box uses obsei ask on the server.
Static export
Section titled “Static export”obsei studio --out site/ # index.html, app.js, styles.css and data.jsonHost the folder anywhere (an internal bucket, Cloudflare Pages behind Access). The export holds only k-anonymous aggregates and a few redacted quotes per theme.
Create a Slack app with a slash command /obsei pointing to
https://your-host/slack/commands, set SLACK_SIGNING_SECRET, and allow hooks.slack.com in the
egress policy. Answers are posted to the channel with record-id citations.