Configure your venue agent's memory and knowledge

In your venue admin, go to Knowledge → Documents and click 'Upload'. Drag in your menu PDF, wine list, or cocktail menu. AGNT's ingestion pipeline splits each document into overlapping chunks (512 tokens, 64-token overlap), runs them through OpenAI's text-embedding-3-small model, and writes them to the VenueChunk table.

Each chunk is tagged by type — menu_item, beverage, dessert — so retrieval can bias per question type. If your menu changes seasonally, re-upload the new version; AGNT replaces the old chunks and re-embeds automatically. The venue_soul cache (Redis, 1-hour TTL) invalidates on every upload.

Supported formats: PDF, DOCX, TXT, CSV
Max file size: 10 MB per document
Chunk size: 512 tokens with 64-token overlap
Embedding model: text-embedding-3-small (1536 dimensions)

Go to Knowledge → FAQs and add question-answer pairs. Start with the ten questions you get most often: hours, parking, dress code, corkage policy, kids' menu, allergy handling, group size limits, reservation policy, cancellation window, and private dining.

Each FAQ entry is embedded separately and tagged as type 'faq'. When a guest asks a question that semantically matches a FAQ, the agent pulls the exact answer instead of generating from general knowledge. This is the single biggest accuracy lever — a venue with 20+ FAQ entries handles 95% of queries correctly on day one.

Under Knowledge → Hours & Rules, set your weekly schedule, holiday closures, and booking constraints. Booking rules include: maximum party size, minimum advance booking time, slot duration, buffer between seatings, and whether walk-ins are accepted.

These rules are enforced at booking time — if a guest asks for a table of 12 but your max is 10, the agent suggests splitting into two tables or escalates to your inbox. Rules are stored as structured JSON on the venue row, not as embedded text, so they are deterministic — no LLM hallucination on hours or capacity.

{
  "max_party_size": 10,
  "min_advance_hours": 2,
  "slot_duration_minutes": 90,
  "buffer_minutes": 30,
  "walk_ins": true,
  "holiday_closures": ["2026-12-25", "2026-01-01"]
}

If your venue handles dietary needs (vegan, gluten-free, halal, kosher, nut-free), go to Knowledge → Dietary and tag each menu item with its applicable restrictions. The agent uses these tags to answer 'Do you have vegan options?' with a concrete list instead of a vague 'Please ask your server'.

You can bulk-import restrictions from a CSV with columns: item_name, restriction_tags (comma-separated). The importer matches items by fuzzy name against your uploaded menu chunks and writes the tags to the VenueMenuItem table.

The venue soul's voice paragraph controls how your agent sounds. Navigate to Soul → Voice and edit the auto-generated paragraph. Good venue voices are specific: 'We are a relaxed beachfront bar — barefoot is fine, reservations are for sunset tables only, everything else is first-come' is better than 'We are a friendly venue'.

You can also set a language priority list. AGNT supports 16+ languages out of the box, but if 80% of your guests speak English and Indonesian, set those as primary so the agent defaults to them when language detection is ambiguous.

Test the voice by sending sample questions in the staging inbox. Adjust the paragraph until the tone matches how your best front-of-house staff member would actually respond.

AGNT provides a verification panel at Knowledge → Health. It runs your top-20 FAQ questions against the current knowledge pack and shows the retrieved chunks, the generated answer, and a confidence score for each. Any answer below 0.7 confidence is flagged.

Walk through each flagged answer. The fix is almost always one of three things: a missing FAQ entry, an ambiguous menu item name (e.g., 'The Special' without context), or a policy that was stated on the website but not uploaded. Fix, re-run the health check, and aim for 90%+ confidence across all 20 questions before going live.