Skip to content

v1.5.0 — Intelligence Layer

Choose a tag to compare

@cpz cpz released this 11 Mar 19:49
· 91 commits to master since this release
18ede2b

Intelligence Layer

Major release adding search intelligence, graph algorithms, code scanning, and session tools.

Search Intelligence

  • Cross-encoder reranker — ONNX ms-marco-MiniLM-L6-v2 rescores hybrid search results by (query, document) pair relevance
  • HyDE — Hypothetical Document Embeddings via LLM for natural language → code search with tree-sitter AST code detection
  • Late chunking — long-context embedding with per-AST-chunk pooling for context-aware chunks
  • Hierarchical summaries (RAPTOR-style) — bottom-up LLM summarization at file → directory → module → repo zoom levels
  • Retrieval feedback loop — tracks search result usage, adjusts future ranking
  • Code-specific embedding models — jina-embeddings-v2-base-code (8192 ctx), nomic-embed-code (32768 ctx), BAAI/bge-m3

Graph & Knowledge

  • Personalized PageRank (PPR) — directed-edge retrieval with correct graph semantics
  • DirectedGraph — preserves edge direction for accurate path/centrality computation
  • Graph algorithms — Tarjan's SCC, articulation points, Dijkstra shortest path, betweenness centrality, Louvain communities with LLM summaries
  • Bridge edges — cross-graph edges connecting code graph and memory graph
  • Co-change analysis — git log mining for files frequently changed together

Code Intelligence

  • Code scanning — succ_index action="scan" with .succignore, symlink rejection, size filtering
  • LSP client — headless JSON-RPC LSP connection for definition, references, hover (Kotlin + Swift servers)
  • MCP review tool — succ_review for code review with blast-radius estimation
  • Repo map — tree-style project structure maps for LLM context

Session & Memory

  • Session surgeon — analyzer (token breakdown), trimmer (tool content, thinking, images), compact with UUID chain integrity
  • Auto-memory extraction — session-end fact extraction with dimension-bucketed consolidation
  • Observability pipeline — search latency, embedding times, LLM call metrics
  • Cross-repo search — search across multiple succ-indexed repositories

Infrastructure

  • API versioning — /v1/ route prefix aliases for all daemon endpoints
  • PostgreSQL CHECK constraints — defense-in-depth for confidence, source_type, link weights
  • succ init improvements — always refresh hooks on rerun, warn on stale project LLM overrides
  • Input normalization — reranker topK/weight clamping, cosine similarity validation, HyDE blank query guard
  • Security — symlink rejection, ReDoS-safe diff parsing, execFileSync for command injection prevention

Stats

  • 148 files changed, +19,241 / −2,677 lines
  • 2,038 tests passing
  • CI green on Ubuntu/macOS/Windows × Node 20/22/24

Full Changelog: v1.4.59...v1.5.0