Cognica provides a PostgreSQL‑compatible engine that combines relational SQL, full‑text search, vector similarity, and graph queries in a single runtime. Its Unified Query Algebra reduces all workloads to a posting‑list abstraction, enabling sub‑millisecond JIT compilation and 2–10× faster queries on up to 100 TB of data. The product can be deployed as a server via the PostgreSQL wire protocol or as an embedded library, offering ACID guarantees across all data modalities.
Funding
Funding not disclosed
Founders
Product
Problem
Enterprises and developers often have to stitch together separate systems—relational databases, search indexes, vector similarity engines, and graph stores—to support diverse data workloads. Managing multiple adapters and data pipelines increases operational complexity, latency, and cost, while preventing unified query optimization across these paradigms.
Solution
Cognica delivers a PostgreSQL‑compatible transactional engine that unifies relational SQL, full‑text search, vector similarity, and graph queries within a single runtime. The core Unified Query Algebra (UQA) reduces every workload to a posting‑list abstraction, enabling Boolean‑algebra based optimization across all query types. Built on modern C++23, the engine offers sub‑millisecond JIT compilation, 2–10× speed improvements, and single‑node capacity up to 100 TB. It supports both server‑mode deployment via the PostgreSQL wire protocol and an embedded library mode for zero‑network overhead. By consolidating these capabilities, Cognica eliminates the need for separate adapters, simplifies architecture, and provides consistent ACID guarantees across all data modalities.
Target Audience
Primary customers are data‑intensive application developers, AI/ML platform teams, and enterprise IT groups that require a single, high‑performance database for relational, search, vector, and graph workloads.
Features
- PostgreSQL wire‑protocol compatibility with full transactional SQL (joins, CTEs, window functions, aggregations)
- BM25 and Bayesian BM25 full‑text search with WAND/BMW optimization, 13 tokenizers, 15 filters, spell‑check, and autocomplete
- HNSW‑based approximate nearest‑neighbor vector search supporting cosine, inner‑product, and Euclidean metrics plus probabilistic score calibration
- Apache AGE‑compatible graph traversal via SQL table functions (BFS, DFS, shortest‑path, pattern matching)
- Unified Query Algebra (UQA) that compiles all paradigms to a single posting‑list structure; includes deep_fusion() for neural‑network expression
- Bayesian BM25 probabilistic scoring enabling log‑odds fusion of lexical, vector, and neural signals
- Flexible deployment: standalone server mode or embedded library mode (SQLite/DuckDB‑style) with <1 ms JIT compile time
- Production‑grade codebase: 800 K+ lines of C++23, 2–10× query speedup, 100 TB single‑node storage capacity