Skip to content

Prism — Overview

Audience: Business stakeholders, product owners, analysts, new team members. This document answers "what does Prism do and why does it matter?" in plain language.

What This Module Does

Prism is a code intelligence service that gives AI coding assistants an accurate, always-current understanding of your codebase. When an AI agent needs to find a function, understand how a module is structured, or know what changed recently, Prism answers the question directly in milliseconds — rather than having the agent read through dozens of files on its own.

The core problem Prism solves is that large codebases do not fit in an AI's memory. A large project might have hundreds of thousands of lines of code across dozens of services and multiple programming languages. Without Prism, an AI agent has to explore blindly, burning time and money reading irrelevant code until it finds what it needs — or gives up entirely. Prism replaces that guesswork with a precise semantic search that understands the meaning of code across 48 programming languages.

Teams connect their repositories via the Prism Console, and Prism automatically builds and maintains a live semantic index. Every push to GitHub updates the index so the AI always works from current code, not a snapshot that is hours or days old. Developers generate a personal token in the Console, drop a small config file into their repo, and their AI assistant immediately gains access to Prism's tools.

Who It Serves

  • Engineering teams using AI coding assistants (Cursor, Claude Code, Windsurf) who want those assistants to navigate large codebases accurately without burning excessive credits on file exploration.
  • Individual developers who want faster, more accurate AI suggestions on large monorepos or multi-service projects where context is hard to provide manually.
  • Team leads and architects who want the AI to understand the full system structure — not just the file currently open — when reviewing designs or planning changes.
  • DevOps and platform teams who need to ensure the AI's understanding of the codebase stays current across many developers working simultaneously.

Key Capabilities

  • Semantic code search across the entire codebase. Ask for "payment webhook validation logic" and get exactly those functions — not a keyword match, but a search that understands what the code does. Works across Python, TypeScript, Go, Java, Rust, and 43 other languages simultaneously. Results are re-ranked by a Google Semantic Reranker for maximum relevance.
  • Structural module maps ranked by importance. For any directory or service, Prism generates a PageRank-scored map showing all modules and their relative importance. The AI gets orientation before diving into unfamiliar code.
  • Go-to-definition for any symbol. When the AI knows the name of a function or class, Prism returns the exact file, line number, signature, and docstring in a single call — no file scanning needed.
  • Reference tracing across the codebase. Find every place a symbol is used, grouped by type: definition, imports, assignments, and calls. Supports batch lookups for up to 10 symbols at once.
  • Always-current indexing that updates on every push. GitHub push webhooks trigger automatic re-indexing. The AI never works from a stale snapshot.
  • Works with any MCP-compatible AI assistant. Cursor, Claude Code, Windsurf, and any other tool that supports the Model Context Protocol can connect to Prism with a single configuration file.
  • Developer tokens that never expire. Each developer generates a personal token in the Prism Console. Tokens are permanent — no hourly refresh cycles or OAuth flows needed.
  • Per-repo scoping. In organizations with multiple repositories, each repo's config file scopes all queries to that repo automatically. The AI never accidentally searches the wrong codebase.

How It Fits in the Platform

Prism is a standalone product in the Swisper ecosystem. It shares GCP infrastructure with Swisper and Swisper Studio but operates independently with its own deployment, authentication, and client base.

From the AI assistant's perspective: Prism is an MCP server — a tool provider. The AI connects to Prism the same way it connects to any other MCP tool, and Prism hands it code search, module maps, symbol definitions, and reference lookups on demand. Prism does not change how the AI reasons or generates code; it changes what context the AI has available.

From the platform's perspective: Prism runs on Google Cloud Run in europe-west1, uses Cloud SQL PostgreSQL with pgvector for storage, and Vertex AI for embeddings. It uses per-tenant schema isolation so each organization's code index is completely separate.

The Prism Console is the web interface where teams manage their Prism account: connect repositories, generate developer tokens, monitor indexing health, and manage team members. The Console communicates with the Prism Gateway API.

Limits and Edge Cases

  • Initial indexing of a new repository takes several minutes. When a repository is first connected, Prism must index the entire codebase. For a 100,000-line codebase this takes approximately 4 minutes; for 500,000 lines, approximately 15 minutes. The index is searchable with partial results while indexing is in progress.
  • The index reflects pushed branches, not local uncommitted work. Changes that exist only on a developer's machine (not pushed to GitHub) are not visible in the hosted index. The index updates when code is pushed.
  • Structural module maps update only on push. The PageRank-based module map is expensive to compute and is only recalculated when code is pushed to GitHub. It reflects the pushed state of the repository.
  • Private repositories require GitHub App installation. For Prism's indexing service to clone a private repository, the Prism GitHub App must be installed on the repository. Public repositories work without additional setup.
  • Currently in early access. Multi-tenant SaaS onboarding is functional, but billing integration and self-hosted enterprise deployment are planned for later phases.

FAQ

Q: What does "40–70% token reduction" mean in practice? A: AI coding assistants charge by the number of tokens they process. Without Prism, an AI navigating a large codebase might read 30 or 40 files before finding what it needs, each read consuming tokens. With Prism, the same task uses a single precise search query. In benchmark testing, this translates to roughly 40–70% fewer tokens consumed per coding task, which directly reduces API costs and speeds up responses.

Q: How is Prism different from GitHub search or Sourcegraph? A: Prism is designed specifically as infrastructure for AI agents, not for human developers. Human code search tools are optimized for how humans think about code (by filename, by keyword). Prism is optimized for how AI agents need context — ranked by semantic relevance, enriched with structural relationships, returned in a compact format the AI can reason about directly. Prism's search combines four retrieval methods (vector similarity, BM25 keyword, description matching, and exact text) fused with Reciprocal Rank Fusion and then re-ranked by a Google Semantic Reranker.

Q: Can Prism index repositories outside our organization? A: Prism can index any Git repository accessible via GitHub. Connect the repository in the Prism Console, install the GitHub webhook, and push once to trigger the first full index.

Q: Does Prism store our source code? A: Prism stores code chunks (small excerpts of code with their embeddings) in Cloud SQL in the europe-west1 region. It does not store complete source files. The indexing worker clones the repository temporarily during indexing and discards the clone after indexing is complete.

Q: What happens if the Prism gateway is down? A: AI assistants will be unable to use Prism tools until the gateway recovers. The gateway runs on Cloud Run with automatic scaling and health checks. Downtime is typically limited to deployment windows (seconds) or rare infrastructure issues.

Q: Which AI assistants work with Prism? A: Any AI assistant that supports the Model Context Protocol (MCP). Confirmed clients include Cursor, Claude Code (Anthropic), and Windsurf. The connection requires adding a single configuration file to your repository.

Q: How do I set up Prism for my team? A: See the Prism Setup Guide for step-by-step instructions. The process takes about 5 minutes: generate a token in the Console, install the CLI (or merge an MCP config for chatbots), and optionally add AI rules so your assistant uses Prism automatically.