Deploying Air-Gapped Llama 3 on Private Enterprise Hardware with vLLM & pgvector

-- Enable the vector extension in PostgreSQL 16+
CREATE EXTENSION IF NOT EXISTS vector;

CREATE TABLE corporate_knowledge_chunks (
    id UUID PRIMARY KEY DEFAULT gen_random_uuid(),
    document_id UUID NOT NULL REFERENCES corporate_documents(id) ON DELETE CASCADE,
    department TEXT NOT NULL CHECK (department IN ('legal', 'finance', 'engineering', 'operations')),
    chunk_index INT NOT NULL,
    chunk_text TEXT NOT NULL,
    token_count INT NOT NULL,
    -- 1536-dim embedding generated locally via BGE-M3 or nomic-embed-text
    embedding vector(1536) NOT NULL,
    created_at TIMESTAMPTZ DEFAULT clock_timestamp()
);

-- Fast Approximate Nearest Neighbor (HNSW) index for sub-10ms queries
CREATE INDEX idx_knowledge_hnsw_cosine 
ON corporate_knowledge_chunks 
USING hnsw (embedding vector_cosine_ops)
WITH (m = 16, ef_construction = 64);

-- Hybrid search function combining full-text search with vector similarity
CREATE OR REPLACE FUNCTION match_enterprise_docs(
    query_embedding vector(1536),
    query_text TEXT,
    dept_filter TEXT,
    match_limit INT
)
RETURNS TABLE (
    chunk_id UUID,
    chunk_text TEXT,
    combined_score FLOAT
) AS $$
BEGIN
    RETURN QUERY
    SELECT 
        c.id,
        c.chunk_text,
        (0.7 * (1 - (c.embedding <=> query_embedding)) + 
         0.3 * ts_rank_cd(to_tsvector('english', c.chunk_text), plainto_tsquery('english', query_text))) AS combined_score
    FROM corporate_knowledge_chunks c
    WHERE c.department = dept_filter
    ORDER BY combined_score DESC
    LIMIT match_limit;
END;
$$ LANGUAGE plpgsql;