Self-improving retrieval, built for production

The knowledge enginethat gets smarter with every query

Cognesys turns your documents into grounded, cited answers — fusing vector, keyword, and knowledge-graph retrieval, scoring every response across five dimensions, and running A/B experiments that promote the winners automatically. Bring your own models.

Bring your own keys Encrypted at rest Per-user isolation No credit card

Cognesys · answer

Reciprocal-rank fusion combines the dense and keyword result lists by summing1the reciprocals of each document’s rank, so passages strong in either signal rise to the top2.

1rrf.md — Hybrid fusion overview
2retrieval/architecture.pdf — Ranking signals
Relevance94
Citation88
Quality91
Grounding96
6
Model providers
5
Evaluation dimensions
3-way
Vector · keyword · graph
Auto
A/B self-improvement
Capabilities

Everything a serious RAG stack needs

Each layer is observable, swappable, and scoped to your data.

Hybrid retrieval

Dense vectors (Qdrant) fused with BM25 keyword search via reciprocal-rank fusion — semantic recall and exact-match precision in a single query.

Cross-encoder reranking

A neural reranker re-scores the top candidates so the most relevant passages reach the model — not just the closest embeddings.

GraphRAG

Build knowledge graphs from your documents, then let entity and relationship links surface connected context that pure vector search misses.

Conversational chat

Multi-turn chat that rewrites each follow-up against the conversation, so questions like “what about its pricing?” just work.

Five-dimension evaluation

Every answer is scored for relevance, hallucination, citation, quality, and context — automatically, in parallel, on every response.

A self-improving flywheel

Evaluation trends spin up A/B experiments and promote the winning configuration on their own. Your retrieval stack tunes itself.

Bring your own models

OpenAI, Anthropic, Google Gemini, Azure, Cohere, or local Ollama — for both LLMs and embeddings. Keys are encrypted at rest, scoped to your account.

How it works

A pipeline that closes the loop

Every query flows through these stages — and every answer feeds the next improvement cycle.

01Rewrite
02Hybrid + graph retrieve
03Rerank
04Generate
05Evaluate
06Optimize
Evaluations drive A/B experiments; winning configurations are promoted automatically — no redeploy.
Works with the models you already use
OpenAIAnthropicGoogle GeminiAzure OpenAICohereOllama

Switch providers per user from Settings — no redeploy. Your API keys are Fernet-encrypted and used only for your own queries.

Polyglot persistence
PostgreSQL + pgvectorQdrantMongoDBRedis

Turn your documents into a knowledge base that learns

Create an account, connect a model, and upload your docs. Cognesys retrieves, evaluates, and improves — automatically.