Dashboard/PRISM/Context/Architecture-Overview.md
PRISM

Context/Architecture-Overview.md

Archived 2026-04-14 — This document describes the v1 architecture (ModuleInterface pattern) which has been fully removed. The current architecture is documented in Architecture/unified-module-architecture.md.

created: 2026-04-08 updated: 2026-04-08 type: context project: PRISM tags: [prism, architecture, context]

PRISM Architecture Overview

Quick-reference architecture context for Claude Code sessions working on PRISM.

What PRISM Is

PRISM (Prospect Intelligence Platform) is an AI-powered system that takes a single domain name and produces consulting-grade competitive intelligence for Algolia sales. It is the SaaS evolution of the Algolia Search Audit skills.

One domain in. 20 modules execute. Scored audit, sales plays, campaign templates, and a factchecked report come out.

Technology Stack

Layer Technology
Backend Python 3.12+, FastAPI, async SQLAlchemy
Orchestration Temporal.io (Python SDK)
Database PostgreSQL 16 (asyncpg driver)
Cache Redis 7 (SSE pub/sub + response cache)
LLM Multi-provider: Anthropic Claude, Google Gemini, OpenAI, OpenRouter
LLM Structuring Instructor library (Pydantic auto-validation + retry)
Frontend Next.js 15, React 19, Tailwind CSS v4, shadcn/ui, Zustand, Vercel AI SDK
Auth Clerk.js

6-Wave Execution Model

Wave 1: Intelligence Collection (13 modules, parallel, ~15 min)
  Gate: intel-company must succeed or audit aborts
Wave 2: Browser Testing (1 module, Playwright, ~10 min, non-fatal)
Wave 3: Factcheck (1 module, child workflow, GAN-inspired)
Wave 4: Insights (1 module, fire-and-forget, benchmarking)
Wave 5: Synthesis (3 modules, parallel, ~15 min)
Wave 6: Final Report (1 module, ~20 min)

Audit modes: full, quick, bulk_triage, refresh.

Module Architecture

Every module implements ModuleInterface with 3 methods: - execute(context) -> ModuleResult - validate(result) -> ValidationResult - health_check() -> bool

Every module follows the file structure:

modules/{module_name}/
  schemas.py     -- Pydantic input/output models
  collector.py   -- External API data collection
  enricher.py    -- LLM structuring (or parser.py if deterministic)
  validator.py   -- 8-10 quality checks
  module.py      -- ModuleInterface implementation + registration

Database Schema (10 tables)

Core: accounts, audits, module_executions, deliverables, vertical_benchmarks Evidence: algolia_customers (2,013), algolia_case_studies (154), algolia_quotes (352), algolia_proofpoints (81), algolia_advocates (268)

API Endpoints

Route Purpose
POST /api/v1/audits/ Start new audit (triggers Temporal workflow)
GET /api/v1/audits/{id} Get audit status + results
GET /api/v1/audits/{id}/stream SSE real-time progress
GET /api/v1/accounts/ List all accounts
GET /api/v1/accounts/{domain}/freshness Check module staleness
POST /api/v1/modules/{name}/execute Execute single module
GET /api/v1/evidence/* Query Algolia customer evidence DB
GET /health Liveness check

External APIs

API Used By Cost
Perplexity (sonar/sonar-pro) 8 intel modules $0.25-3/M tokens
BuiltWith intel-techstack API key tier
SimilarWeb intel-traffic API key tier
Yahoo Finance (yfinance) intel-financial-public Free
SEC EDGAR intel-financial-public, intel-investor Free
Apify intel-hiring, intel-social Credits
Crossbeam intel-partner Deferred (Perplexity fallback)

Compliance Status (vs Vault Standards)

As of 2026-04-08. See Compliance-Assessment for details.

  • CodingSOPs: Mostly compliant. Gap: no adapter pattern (DI), function length.
  • TestingSOPs: Partial. Gap: no TDD, no adapter-based mocks, flat test structure.
  • FolderStructure: Non-compliant (expected debt, predates standard).
  • Manifesto: Core spirit compliant. Gap: no shared-py, no persist(), dict output.