AI Integration

Configure AI chat, LaTeX error suggestions, quotas and search services on a Kubernetes deployment.

The AI assistant and the LaTeX error assistant run in the web service (the workbench and error-assistant modules) of every Ayakaleaf Pro pod. What each setting does, user access and quotas are described in AI Integration; this page puts it into Kubernetes.

Service	web, in every Ayakaleaf Pro pod; read when web starts. The web_fetch tool goes through linked-url-proxy in the same pod
Pod → outside	AI_BASE_URL; WEB_SEARCH_URL (default https://api.tavily.com/search) and DOCS_MCP_URL if search is on; pages the model reads with web_fetch
Outside → Overleaf	none
State	the token quota counters in Redis, shared by all replicas; the chat history stays in the browser
Periodic jobs	none

Configuration

1. ConfigMap

   30-config.yaml (additions)

     AI_ENABLED: "true"
     AI_BASE_URL: https://ai-gateway.example.com/v1
     AI_MODEL: YOUR_TEXT_MODEL_ID
     # optional, see the Toolkit page
     # AI_IMAGE_MODEL: YOUR_IMAGE_CAPABLE_MODEL_ID
     # AI_PDF_MODEL: YOUR_PDF_CAPABLE_MODEL_ID
     # AI_TOKEN_QUOTA: "200000"
     # AI_TOKEN_QUOTA_PERIOD: month
     # AI_PROXY_URL: http://proxy.example.com:3128
   

   AI_ENABLED must be exactly "true". The assistant appears only when AI_ENABLED, AI_BASE_URL, AI_MODEL and AI_API_KEY are all set.

2. Secret

   kubectl -n overleaf patch secret overleaf-secrets --type merge -p '{"stringData":{
     "AI_API_KEY":"<gateway key>",
     "TAVILY_API_KEY":"<search key, optional>"}}'
   
3. Apply, restart and check the gateway from every pod

   kubectl apply -f 30-config.yaml
   kubectl -n overleaf rollout restart deployment/overleaf     # or statefulset/overleaf
   for p in $(kubectl -n overleaf get pod -l app=overleaf -o name); do   # app=overleaf-ha for several replicas
     kubectl -n overleaf exec "$p" -- curl -sS -o /dev/null -w "$p %{http_code}\n" https://ai-gateway.example.com/v1/models
   done
   

   Any HTTP status means the gateway is reachable; 000 means it is not.

The web_fetch tool fetches pages through the same proxy as External URL files, so the OVERLEAF_LINKED_URL_* settings also limit what the model may read.

Several replicas

Nothing extra. The token quota is counted in Redis, so a user’s usage is the same on every replica, and an admin’s reset applies everywhere.

Tested on Kubernetes

Against a mock of an OpenAI-compatible gateway, Tavily and the GitBook MCP endpoint (trusted through NODE_EXTRA_CA_CERTS from overleaf-certs), on Kubernetes 1.37, our end-to-end tests pass: for the AI assistant on three replicas behind HAProxy (31 cases), chats with tool calls and suggested edits, web and documentation search, reading web pages through linked-url-proxy limited by OVERLEAF_LINKED_URL_ALLOWED_RESOURCES, image and PDF attachments, consent, the token quota counted in Redis, AI turned off for a user and for the instance, and slow connections to the gateway; for the error assistant on one replica (15 cases), suggested fixes, consent, the shared token quota and AI turned off.