🧠 Semantic memory
Passages embedded with fourth-generation multilingual embeddings alongside a custom Arabic lexical index. Retrieval is hybrid: meaning and term fused by reciprocal rank, then reranked — capturing both the modern question and the classical term.🕸️ Knowledge graph
Concepts, scholars and relations (grounds, critiques, explains, branches from…) are extracted from passages, and every relation carries a verbatim supporting quote with its page — the whole graph is human-auditable, thread by thread.⬜ Principle: empty is valid
The extractor is forbidden to infer beyond the text; a passage with no explicit relation remains relation-free. Precision over completeness.The research connector
The corpus is served over the Model Context Protocol with OAuth 2.1 sign-in and a governed access list. All tools are read-only: hybrid search, graph exploration, thematic communities, book card, sequential reading, and service description.
For the technically curious
Ingestion stages — semantic memory
- Structural parsing — a dedicated parser for the Library’s book format merges footnotes at their anchors and rebuilds hierarchy; a unifier does the same for word-processor files. [[PAGE_NO]]
- Dense embedding — one multilingual, multimodal vector per passage. Cohere embed-v4.01536-dim
- Lexical indexing — sparse vectors from a custom Arabic analyzer for exact term matching. BM25-style sparse
- Hybrid retrieval — both arms fused in a single unified-payload vector collection. QdrantRRF
- Reranking — a fourth-generation reranker lifts the most relevant to the top. rerank-v4.0
- Quality gate — query caching and calibrated relevance thresholds prevent weak results being served as answers. Redis
Knowledge graph & archival
- Graph — book, passage and concept nodes with evidence-bearing edges, full Arabic text analysis. Neo4jGDS
- Versioning gate — SHA-256 per file: deduplication, atomic supersession of newer editions, text-over-scan precedence.
- Archive & backup — a permanent object archive for originals and automated nightly snapshots of the vector store. MinIO / S3
Memory answers: what was said about X? The graph answers: how does X relate to Y? Together they cover the direct question, the relational question, and the map of a subject.