MaaCare Docs
Sign inCreate account

Algorithms & behavior

This page describes non-secret logic that shapes product behavior. For file-level references, see the linked source modules.

Chat reply language

Before RAG, the chat route runs prepareMultilingualChatTurn (src/lib/chat/multilingual-prep.ts): one LLM call returns strict JSON (validated with Zod) containing an IETF language tag (BCP-47) for the user’s latest message, an English retrieval query for embedding over the English-only corpus, and an optional human-readable language hint for the answer model. On parse failure it falls back to English and the raw user text for retrieval.

The final model call is instructed to reply in that language while retrieved passages stay in English. Users may still mix languages in conversation; the latest turn drives detection.

Transcript budgeting

Long conversations are trimmed with a token budget heuristic (character length ÷ 4) so prompts stay within model limits. Older turns may be summarized or dropped before the model call.

Retrieval-augmented generation (RAG)

  1. User (or server) issues a query relevant to medical knowledge.
  2. A multilingual prep step produces an English retrieval query (any source language → concise English for semantic search).
  3. Embedding via the configured Gemini embedding model.
  4. Vector search in Postgres (rag_chunks) through a SQL RPC matcher.
  5. Retrieved passages are injected into the system or tool context with citations-style provenance where implemented.

LLM failover

Primary text generation uses Google Gemini. On failure or empty output, the server falls back to Groq (OpenAI-compatible API) with the same prompt envelope where possible.

Nearby facilities (Bangladesh)

When the client passes approximate coordinates and the user intent matches nearby care, the server can inject Bangladesh facility catalog rows into chat context to ground facility answers.

Community realtime (optional)

When Supabase Realtime publication includes community tables, the client may:

  • Subscribe to postgres_changes for comments, likes, or posts.
  • Debounce rapid bursts of events before refetching lists to avoid UI thrash.

Policies still apply: clients only receive events for rows they could read with their JWT.

Notifications

Row inserts from community activity (for example new comments or likes) can enqueue notifications rows via database triggers. The bell UI merges REST polling with optional realtime subscriptions.

Mermaid diagram