Frequently asked questions

Benchmark snapshot

We run an internal 30-case evaluation suite covering legal-statutory and coaching prompts on every model and prompt change. Results below are from the most recent production-endpoint run.

Mixed v2 suite

23/30 cases passed - legal-statutory + coaching prompts

Coaching mode

12/15 women-coach cases

Legal mode

11/15 statutory + ambiguous-legal cases

Each case has expected patterns (citations, articles, refusals). Eval scripts hit the live API, score against expected outputs, and emit a markdown report.

About AMSA

What is AMSA?

AMSA is a civic AI mentor focused on Moroccan local governance - communes, advisory bodies, petitions, council preparation, and gender-sensitive participation. It answers in Arabic and Moroccan Darija using a curated corpus of official Moroccan sources (Constitution, Organic Laws 111/112/113.14, DGCT guides, IEECAG materials).

Who is it for?

Elected women, civil society leaders, advisory-body members, youth-organization staff, and citizens preparing to engage with their commune. AMSA is designed for non-technical users - no legal background required.

Is it free?

Yes. AMSA is free for end users. A daily fair-use quota prevents abuse but resets automatically every morning.

Accuracy & limits

How accurate is AMSA?

On our 30-case mixed evaluation, AMSA passes 77% - with 80% accuracy in coaching mode and 73% in legal mode. See the benchmark table above for the breakdown.

What does AMSA refuse to do?

AMSA never gives binding legal advice, never represents itself as a lawyer or public authority, and refuses to fabricate citations. When a question falls outside its corpus or requires a human expert, it says so and points the user toward the right institutional channel.

Can it hallucinate?

Like any LLM, yes - but every answer is grounded in retrieved corpus chunks with article-level citations. Users can click any citation to see the underlying text. We also track hallucination-trap cases in our benchmark and currently catch 100% of them.

Languages & accessibility

Does AMSA understand Moroccan Darija?

Yes - that's a core design goal, not a translation afterthought. We detect Darija inputs, rewrite them to Modern Standard Arabic for retrieval, and answer in the user's original register.

Is there an English version?

The public marketing site is bilingual (Arabic + English). The chatbot itself answers in Arabic / Darija because its corpus and user base are Moroccan.

Privacy & safeguards

What data is stored?

Conversations are stored (anonymized by an opaque user ID) so we can improve answer quality and run evals. No phone numbers, real names, or identifying personal data are required to use AMSA.

Where does AMSA run?

On secure managed cloud infrastructure - globally distributed - with the corpus and conversation history in private databases. No data leaves our infrastructure except for the language model provider call needed to generate answers.

How do I report a bad answer?

Every assistant message has a 'Report' button (the flag icon). Reports go straight to the project Slack for review. You can also use the 'Report / Propose' button in the top navigation for general feedback.