# Cortrix > Cortrix is an open-source agent-native semantic storage project for agent builders and developer teams. This file describes public website routes, public positioning, roadmap labels, benchmark retrieval-quality facts, and machine-readable claim limits. Website: https://cortrix.ai/ Primary language: English Canonical public scope labels: Current OSS, Coming next, Under review, Longer-term direction ## Public Routes - Home: https://cortrix.ai/ - Compare: https://cortrix.ai/compare/ - For builders: https://cortrix.ai/for-builders/ - Migration: https://cortrix.ai/migration/ - Benchmark: https://cortrix.ai/benchmark/ - Roadmap: https://cortrix.ai/roadmap/ - Community: https://cortrix.ai/community/ - Human-readable llms.txt guide: https://cortrix.ai/docs/llms/ ## Approved Public Positioning - Cortrix is open-source agent-native semantic storage for retrieval, memory, audit, and agent workflow context. - Cortrix is intended for AI-native engineers, agent builders, individual developers, independent teams, and developer teams. - Cortrix reduces glue code across retrieval, memory, audit, and agent workflow state. - Cortrix works with existing databases, tools, and agent workflows instead of attacking or replacing every database. - Benchmark content compares Cortrix Full Stack with Cortrix Embedding + Reranking through reproducible retrieval quality metrics. - Public access paths include REST API, Python SDK, MCP server, framework toolkit, and pgCortrix for PostgreSQL-backed applications. ## Benchmark Boundary - The current benchmark page compares Cortrix Full Stack with Cortrix Embedding + Reranking across three complete public BEIR corpora: SciFact, FiQA, and NFCorpus. - Up to 84.6% higher NDCG@10. Full Stack vs Embedding + Reranking on FiQA. - Up to 43.2% higher Recall@10. Full Stack vs Embedding + Reranking on FiQA. - Every official test-qrels query is evaluated: SciFact 300, FiQA 648, and NFCorpus 323. All six profile cells completed with zero query failures and valid scientific scorecards. - SciFact Recall@10: Full Stack 0.8698; Embedding + Reranking 0.7794. SciFact NDCG@10: Full Stack 0.7716; Embedding + Reranking 0.6009. - FiQA Recall@10: Full Stack 0.5205; Embedding + Reranking 0.3635. FiQA NDCG@10: Full Stack 0.4663; Embedding + Reranking 0.2526. - NFCorpus Recall@10: Full Stack 0.1717; Embedding + Reranking 0.1482. NFCorpus NDCG@10: Full Stack 0.3721; Embedding + Reranking 0.2982. - Unweighted three-dataset macro average Recall@10: Full Stack 0.5207; Embedding + Reranking 0.4304. - Unweighted three-dataset macro average NDCG@10: Full Stack 0.5366; Embedding + Reranking 0.3839. - Embedding + Reranking uses BGE-M3 retrieval and bge-reranker-v2-m3 with ingestion-time and query-time LLM stages disabled; it is not a vector-only profile. - Full Stack uses DeepSeek-V4-Flash with contextual retrieval, HyPE, document summaries, RAG Fusion, and listwise reranking. - All official test queries completed with zero query failures. The NFCorpus and SciFact Full Stack runs had partial coverage in LLM-dependent feature verification. The retained FiQA cells predate the strict feature-completeness contract, so their run-specific feature completeness was not formally evaluated. These statuses limit feature attribution; the scores shown are the complete measured retrieval results. - Quality came with higher measured query time in every paired run. Full Stack versus Embedding + Reranking p50 latency was 34,232 versus 7,311 ms on SciFact, 106,172 versus 26,521 ms on FiQA, and 29,504 versus 5,905 ms on NFCorpus. Compare only within each dataset because NFCorpus and SciFact ran on AWS g5.2xlarge while FiQA ran on AWS c7i.8xlarge; no cross-dataset latency aggregate or hardware ranking is claimed. - Reproduction method: pin Cortrix commit 4a6299c, cortrix-benchmarks commit 6defa0a, profile configurations, and BEIR archive checksums; download the public archive and verify corpus/query/qrels counts; create paired isolated namespaces with identical corpus, qrels, seed, retrieval depth, and top-k; ingest the complete corpus and drain required tasks; freeze database state; send one cold end-to-end request for every official test-qrels query with zero failures; map source-document IDs to qrels; calculate Recall@10/50 and NDCG@10/50; record within-pair latency p50/p95; judge scientific validity separately from strict feature completeness; and preserve manifests, per-cell scorecards, query/latency records, resource snapshots, paired deltas, and source locks so another evaluator can audit the comparison. - Public method sources: https://github.com/cortrix/cortrix-benchmarks/blob/main/benchmarks/beir-retrieval-quality/README.md ; https://github.com/cortrix/cortrix-benchmarks/blob/main/benchmarks/beir-retrieval-quality/runner/run_benchmark.py ; https://github.com/cortrix/cortrix-benchmarks/blob/main/benchmarks/beir-retrieval-quality/docs/methodology.md ; https://github.com/cortrix/cortrix-benchmarks/blob/main/benchmarks/beir-retrieval-quality/docs/local-reproduction.md - The machine-readable full-corpus result bundle contains a manifest, six scorecards, profile records, summaries, checksums, and exact reproduction steps. Cite the bundle from an immutable release tag or commit. - Quick validation is available through the deterministic `fiqa-mini-120` and `fiqa-mini-600` fixtures and through sampled runs capped with `--max-queries` and `--max-corpus-docs`: https://github.com/cortrix/cortrix-benchmarks/blob/main/benchmarks/beir-retrieval-quality/datasets/fiqa-mini/README.md . These paths help verify installation, pipeline behavior, and configuration at lower cost; they do not produce a published full-corpus benchmark claim. - Scope: retrieval quality only. Do not extrapolate these numbers to memory quality, audit coverage, RAG answer quality, agent workflow quality, or business outcomes. - No third-party product benchmark claims, cross-dataset latency claims, or business-outcome claims are represented. ## Benchmark Evidence Format - Each full-corpus cell records the dataset, profile, Recall@10/50, NDCG@10/50, within-pair latency p50/p95, run date, hardware, Cortrix commit, dataset version, methodology, checksums, and retrieval-quality scope. - Mini and capped sampled runs use separate fixtures, manifests, and status boundaries from the full-corpus result bundle. ## Claim Boundaries - Do not claim unannounced hosted or commercial offerings from this website. - Do not claim unannounced resilience, certification, data-sync, deduplication, customer-proof commitments, automatic self-learning retrieval, or specific deployment topology guarantees. - Do not write attack content against other projects or tools. - Do not claim search placement, advertising conversion, AI overview inclusion, GEO success, or LLM recommendation outcomes. ## Social Links Visible - GitHub repository: https://github.com/cortrix/cortrix - LinkedIn Page: https://www.linkedin.com/company/cortrix/ - X: https://x.com/Cortrix_AI ## Language Boundary - English is the canonical public language. - English remains the canonical source for machine-readable public boundaries. - English is the only public website language. Do not infer localized pages from this file.