sqlite-muninn

Runs GGUF chat models and structured extraction (NER, relation extraction, combined NER+RE, summarization) inside SQLite via muninn_chat(), muninn_extract_entities(), muninn_extract_relations(), muninn_extract_ner_re(), muninn_summarize(), and their _batch variants. Supports supervised (labels provided) and unsupervised (open-extraction) modes, and GBNF grammar-constrained JSON output. Use when the user mentions "muninn_chat", "named entity recognition", "NER", "relation extraction", "RE", "LLM in SQL", "GGUF chat model", "Qwen3.5", "Qwen 3.5", "GBNF grammar", "grammar-constrained", "summarization in SQLite", "structured extraction", or wants to run an instruction-tuned model inside SQLite.

Generates text embeddings inside SQLite via muninn_embed() backed by local GGUF models (BERT, MiniLM, nomic-embed-text, BGE-M3, Qwen3-Embedding) through llama.cpp with Metal GPU acceleration on macOS. Covers muninn_embed_model registration, the temp.muninn_models virtual table, composing with hnsw_index for semantic search, and auto-embed TEMP triggers. Use when the user mentions "text embedding", "semantic search", "sentence embedding", "muninn_embed", "GGUF embedding model", "MiniLM", "nomic-embed", "BGE-M3", "Qwen3 embedding", "embed model in SQLite", "text-in semantic-search-out", or wants to embed text directly in SQL.

Runs graph algorithms (BFS, DFS, shortest path, PageRank, connected components, degree / node-betweenness / edge-betweenness / closeness centrality, Leiden community detection) on any SQLite edge table via table-valued functions. Covers the WHERE edge_table = ... constraint calling convention and the persistent CSR adjacency cache (graph_adjacency). Use when the user mentions "graph traversal", "BFS", "DFS", "shortest path", "Dijkstra", "PageRank", "connected components", "centrality", "betweenness", "closeness", "Leiden community", "community detection", "graph algorithm in SQLite", "graph_bfs", "graph_pagerank", "graph_leiden", "CSR adjacency", "graph_adjacency", or wants to analyze a network / dependency / social graph stored in SQLite.

Writes dbt-inspired lineage queries using the graph_select TVF — ancestors, descendants, depth-limited traversal, transitive closures, and set operations on any SQLite edge table. Selector DSL supports +node (ancestors), node+ (descendants), N+node+M (depth limits), @node (closure), space (union), comma (intersection), and "not" (complement). Use when the user mentions "graph_select", "dbt selector", "lineage query", "dependency graph", "upstream", "downstream", "ancestors", "descendants", "transitive closure", "impact analysis", "build closure", "dead code detection", or wants dbt-style node selection syntax on any SQLite edge table.

Builds end-to-end GraphRAG retrieval over a text corpus entirely in SQLite: chunk → embed → extract entities and relations → build knowledge graph → detect communities → label clusters → retrieve via vector seed + graph expansion + centrality ranking. Composes muninn_embed, hnsw_index, muninn_extract_ner_re_batch, graph_leiden, muninn_label_groups, and muninn_extract_er. Use when the user mentions "GraphRAG", "Graph RAG", "knowledge graph RAG", "retrieval-augmented generation", "RAG pipeline in SQLite", "KG retrieval", "entity resolution", "muninn_extract_er", "community labeling", "KG indexing", or wants to build Microsoft GraphRAG- style retrieval on a corpus.

Trains Node2Vec structural graph embeddings (Grover & Leskovec, 2016) from a SQLite edge table and writes them directly into an hnsw_index virtual table. Covers the node2vec_train scalar function, p/q walk bias tuning, window size, negative sampling, epochs, and composing with KNN vector search. Use when the user mentions "node2vec", "graph embedding", "structural embedding", "random walk embedding", "DeepWalk", "node2vec_train", "similar nodes", "graph representation learning", or wants to compute embeddings from graph topology (not text).

Installs, loads, and smoke-tests the muninn SQLite extension across SQLite CLI, C, Python, Node.js, and WASM runtimes. Covers pip install sqlite-muninn, npm install sqlite-muninn, prebuilt .dylib/.so/.dll from GitHub Releases, and from-source builds with make. Use when the user mentions "install muninn", "load muninn", "load the extension", "enable_load_extension", ".load ./muninn", "sqlite3_muninn_init", "sqlite-muninn", "pip install sqlite-muninn", "npm install sqlite-muninn", "muninn.dylib", "muninn WASM", or asks to get started with the library.

Diagnoses muninn build failures, extension-loading errors, platform-specific pitfalls, and runtime issues across SQLite CLI, C, Python, Node.js, and WASM. Covers SQLITE_OMIT_LOAD_EXTENSION on macOS system Python, CMake hangs on Apple Silicon (GGML_NATIVE), Metal GPU toggles (MUNINN_GPU_LAYERS), llama.cpp log verbosity (MUNINN_LOG_LEVEL), ASan-built dylibs failing in Python, GGUF model resolution, and wasm32 vs wasm64 linkage. Use when the user mentions "unable to open shared library", "not authorized", "SQLITE_OMIT_LOAD_EXTENSION", "enable_load_extension fails", "muninn build failed", "CMake hangs", "Segmentation fault muninn", "model not registered", "MUNINN_GPU_LAYERS", "WASM muninn fails", "muninn troubleshooting", or hits a muninn load / build error.

Builds and queries HNSW approximate-nearest-neighbor vector indexes in SQLite using the hnsw_index virtual table. Covers CREATE VIRTUAL TABLE with dimensions/metric/m/ef_construction, INSERT of raw float32 blobs, MATCH queries with k and ef_search, and vector blob encoding in Python (struct.pack), Node.js (Float32Array), and C. Use when the user mentions "vector search", "HNSW", "nearest neighbor", "KNN", "similarity search in SQLite", "hnsw_index", "cosine similarity", "L2 distance", "inner product", "vector MATCH", "ef_search", "float32 blob", or asks to add semantic / similarity search to a SQLite database.