Documentation

Resources

Pipeline architecture and working prompts from the Vera-local docs — embedded here for quick reference. Click any prompt to copy it for Vera Chat.

Pipeline

Deterministic flow: prompt → summary

Vera routes every prompt through a fixed, deterministic pipeline — no LLM on the hot path. Agent and query paths share one chat shell, then diverge at the router.

Input

Chat / Send URL ?prompt= Agent chip

Router — agent trigger match?

Agent path

agent_registry → agents/*.js → summarizeAgent

Query path

buildQueryPlan → executePlan → normalize → summarize

Output — prompter.js

Human summary Technical JSON Share / TSV / CSV

Bootstrap loads skills, org profile, catalogs, planner, query engine, and summarizer before any prompt runs. Organization profiles map logical field names to physical columns; vocabulary lives in skills — not in core JS.

Path	When	Core output
Query	No agent trigger matches	Structured query plan → filtered rows, counts, or aggregates
Agent	`agentSkill` trigger matches (or chip forces agent)	`AgentResult` sections — multi-dataset logic, customer-portfolio joins

Prompts

Working prompts by category

Curated examples supported by the regression harness and demo org — click to copy, then paste in Vera Chat.

Amounts, rates & dates

Comparison operators, rolling windows, and multi-constraint filters on loans and deposit products.

Product type (classCode)

Numeric product codes via bankingSkill shared field phrases (of type, class code, …).

Customer ID

Portfolio / customer / member scoping with 6–12 digit identifiers.

Umbrella accounts

Cross-product account lookups by customer id or display name.

Valuation words & idioms

Natural-language amounts — millions, shorthand k, and between phrasing.

Product translator

Named products from curated.products — catalog resolution to class or type codes.

Employee translator

Employee codes and display names via curated.employees — booked by, assigned to, for employee N.

Location translator

Branches, tokens, and US state resolution via curated.locations.

Counting & distinct counts

Row counts and unique classCode values — catalog-defined vs in-portfolio.

Ranking & statistics

Sort + limit and aggregate functions (mean, median, stdev, …).

Version 4 — customer-portfolio agent

Cross-product customer joins and exclusions — handled by depositCustomerSearch, not row-level query plans alone.

Reference

Organization profiles & agents

Orgs

Organization profiles

Each deployment maps logical field names to physical columns in orgs/<orgId>/profile.js. Switch active orgs via orgs/active.js — copy orgs/demo/ as a starting point.
Agents

Vera AI agents

Standalone scripts in agents/ implement run(context, callback) and return AgentResult sections. Triggers live in skills/agentSkill.js; heavy cross-product logic (customer profitability, portfolio search) runs here instead of row-level query plans.

Repository map

Key paths

skills/generalSkill.js     ← cross-domain language + scope polarity

skills/bankingSkill.js    ← banking vocabulary + translators

skills/agentSkill.js      ← agent catalog + triggers

orgs/<orgId>/profile.js  ← your column mappings

js/prompt.js             ← buildQueryPlan

js/query_engine.js      ← executePlan

js/conversation.js      ← multi-turn sessions

js/prompter.js          ← chat UI

agents/*.js             ← heavy computation

scripts/test-*.js       ← regression harness

Tests

Regression harness

Run locally with Node.js to validate prompts, conversation, and agents.

node scripts/test-prompt-regression.js

node scripts/test-conversation.js

node scripts/test-clarification.js