Themes, dedupe and ask
obsei themes # embed new feedback, update themes, list themobsei ask "Why are users in Brazil asking for refunds?"Stable themes
Section titled “Stable themes”Each new record is embedded, marked as a near-duplicate of an earlier record when it is almost identical, and assigned to the nearest theme or starts a new one. Theme ids never change, so dashboards, tickets and agents can refer to them over time.
themes: embedder: auto # local when obsei[embeddings] is installed, else hashing; or an llms name similarity: null # default 0.3 hashing, 0.65 local multilingual, 0.75 other models duplicate_similarity: 0.9 labeler: ollama # llms name; without it, keyword labels label_language: English # labels in your team's language; feedback stays as written k_anonymity: 5 max_records: 10000 # records embedded and assigned per refresh batch_size: 64 # records per embedding callWithout a labeler, a theme is named by up to three words that several of its records share and
that are rare in the rest of the store, so words most feedback uses (stopwords, an issue template’s
headings) are left out. Labels are set once, when a theme first reaches k_anonymity people.
The built-in hashing embedder needs no model and works in every script, but it groups by
wording, so the same issue in two languages forms two themes. To group by meaning across 50+
languages, use the local multilingual model (CPU only, about 220 MB, no PyTorch):
uv tool install --force "obsei[mcp,embeddings]>=1.0" # or: pip install "obsei[embeddings]>=1.0"obsei models download --dir /models # once, with networkexport OBSEI_MODELS_DIR=/models # air-gapped runs load it from hereWith the extra installed, the default embedder: auto uses this model; set embedder: local to
fail instead of falling back to hashing when it is missing:
themes: embedder: local # or local:intfloat/multilingual-e5-large for higher qualityThresholds differ between models; tune similarity if themes split or merge too much. A model
served by Ollama or vLLM (for example bge-m3) also works:
llms: ollama: base_url: http://localhost:11434/v1 model: qwen3:8b embedding_model: bge-m3themes: embedder: ollama # the llms name, not "local"auto, hashing, local and local:<model> always mean the built-in embedders, so do not use them as
llms names for an embedding endpoint.
k-anonymity
Section titled “k-anonymity”Themes from fewer than k_anonymity people are never labelled, listed, exported or shown in
Studio, so a small theme cannot single out a customer. Record counts from feedback_stats and
search results are not filtered by k.
obsei ask retrieves the most relevant feedback, adds the theme overview, and asks your model to
answer in the language of the question, quoting customers in their own language and citing record
ids. Citations are checked against the records actually retrieved.