MiiQ retrieval-intent classifier (multilingual-e5-small)
Classifies a chat query to an enterprise knowledge-base assistant into one of 14 retrieval intents, so the assistant can choose how to search its documents.
| Intent | Meaning |
|---|---|
exact_lookup |
One stated fact, a named clause/section/form/code, or exact wording |
term_search |
A bare keyword, or where a term is mentioned |
compliance_check |
Whether something is allowed, required, eligible or covered |
explanation |
What something means, why, how it works, what a policy says |
comparison |
Differences or similarities between named things |
computation |
Arithmetic: totals, counts, durations, percentages |
reasoning |
What-if, implications, or a judgement/recommendation drawn from documents |
procedural |
Steps or process for doing something |
summarization |
Summary or key points of a document |
greeting, thanks, farewell, chitchat |
Social messages (chitchat is about the assistant itself) |
off_topic |
Unrelated to the organisation's work |
Model
intfloat/multilingual-e5-small, fine-tuned end to end with a linear layer (cross-entropy, 4 epochs). The encoder is stored in half precision; the head is intent_head.json (classes, coef, intercept).
To score a query: prefix it with "query: ", mean-pool the last hidden state over the attention mask, L2-normalise, then apply softmax(coef · x + intercept).
Training data
842 English and Arabic queries written for this purpose from a labelling guide, with AI assistance: workplace questions about HR, finance, IT, procurement, contracts, safety and equipment documents, questions about the user's own uploaded files, social messages and off-topic requests. No user data was used.
Evaluation
| Set | Queries | Accuracy | English / Arabic |
|---|---|---|---|
| Held-out, written separately from the training data | 243 | 93% | 93% / 92% |
| Held-out questions about the user's own files, with contrasts about the assistant | 97 | 88% | 95% / 76% |
| Real queries typed during internal testing | 82 | 80% |
Median latency is about 23 ms per query on four CPU threads.
Limits
- Several intents overlap by nature (for example
exact_lookupandcomputation), and labels follow one labelling guide. - The real-query sample is small.
- The model is tuned for workplace knowledge-base assistants and has not been evaluated elsewhere.