Code intelligence for AI agents and the humans alongside them.

Coding agents

Claude Code, Cursor, Continue, and any editing agent call the CLI directly. Automatic git-overlay is on by default, so unpushed local edits are visible to search and refactor tools before they are committed.

Non-editing agents

JWT-authenticated, per-tenant Postgres-schema-isolated. Wiki agents, architect agents, ChatGPT, Claude Desktop, and custom internal tooling consume it over Streamable HTTP.

Four-leg hybrid retrieval

Most “hybrid” tools fuse vector + BM25. Prism fuses four legs via Reciprocal Rank Fusion (k=60): vector, BM25, description-vector, and exact ILIKE.

Architecture intelligence

PageRank + Leiden clustering means Prism knows which files matter and which files belong together. get_module_map and get_architecture return ranked components, not file lists.

Live unpushed-file overlay

The CLI runs git status, attaches dirty files as an overlay payload, and the gateway re-parses them in-memory before answering. No one else does this without forcing the agent to ship full-file payloads on every call.

Adaptive response shaping

Tools return ~200–400 tokens compact, ~500–3000 with snippets, automatically windowed when results are large. Confidence and provenance are stamped on every response so agents know when to act vs. reformulate.

Reuse-first discipline as a tool

check_exists matches a natural-language intent against LLM descriptions to surface existing helpers before the agent writes a duplicate. prepare_to_edit bundles source, top callers, test files, and rules in one call.

Per-tenant schema isolation

Not row-level filtering, not application-layer auth: physical schema separation as tenant_{id}. One SQL injection in a tenant query cannot escape the tenant.

Shallow git clone into a per-job tempdir, with the working tree wiped before each run.

→ Safe parallel indexing across many tenants on shared infrastructure.

SHA-256 hash per file, diffed against file_index_state. Tags every file as created, modified, deleted, or unchanged.

→ A typical PR-sized push reindexes in seconds, not minutes.

Tree-sitter AST chunking at function, class, and method boundaries. Wave-bounded concurrent embedding via Vertex AI gemini-embedding-001 (2000-dim). COPY bulk store.

→ Function-level search hits instead of random text windows. Large repos index reliably under quota turbulence.

Parses .cursor/rules/*.mdc and .prule/* files at section granularity, then embeds each section.

→ An agent can pull a single rule section without loading the whole 200-line rule file. Saves ~80–90% of the context.

Tree-sitter scope analysis across 11 languages. Extracts every symbol reference and tags it as definition, import, call, or assignment.

→ Refactoring is safe. Find all callers of User.save returns the real blast radius, not noise from imports and prose.

Anthropic Batch API generates a one-sentence summary per chunk, embeds the summary, and stores it as a second embedding column alongside the code embedding.

→ Natural-language queries work as well as exact-keyword queries. search_code("rate limiting sliding window") finds code that doesn't contain those words.

Import-graph extraction across all source languages, then PageRank over the graph. Outputs per-module importance_score (0–100) and tier (core / significant / peripheral).

→ The same algorithm that built Google Search, applied to your codebase. The foundational files automatically rise to the top.

Persists import edges to the database. Two modes: full rebuild on structural change, delta update for content-only edits. The pipeline picks the right mode automatically.

→ Webhook-driven incremental indexing stays fast even on monorepos with thousands of modules.

Leiden clustering over the import graph (modern community-detection algorithm, successor to Louvain). Produces components with LLM-generated summaries, capability index, and divergence detection.

→ Architects see the system as it really behaves, not as the folder tree suggests. Drift is flagged before it degrades the architecture.

UPDATE-JOIN propagating component_id from module_summaries back into code_chunks.

→ Search results know which architectural component they belong to. Agents scope queries to "the auth component" without knowing the folder structure.

PostgreSQL tsvector triggers maintain the BM25 full-text index automatically as data flows in.

→ The exact-keyword and BM25 legs of search are always in sync with the latest pipeline run, with zero operational overhead.

Vector cosine

Code embeddings over tree-sitter AST chunks. Finds code by structural and lexical similarity.

BM25 (tsvector)

PostgreSQL full-text index over identifiers, keywords, and comments. Updated automatically by triggers.

Description vector

LLM-summarized chunks indexed as a second embedding. Lets natural-language queries match what the code does, not what it says.

Exact ILIKE

Strict superset of grep. Guarantees agents never need to fall back to text search.