--- title: "QueryLlama Features" description: "Keyword, semantic and hybrid search, RAG answers with citations, bring-your-own-LLM, and 5-role RBAC enforced by Postgres Row-Level Security." url: "https://royalsoftworks.com/products/queryllama/features/" source: "https://royalsoftworks.com" format: "markdown" note: "Markdown rendering of the HTML page at `url`. Same content, same canonical URL." --- Features # Search, then ask. Always with a source. QueryLlama is being built around one idea: every answer should point back to the document it came from. Here's what that means in practice, and exactly how much of it is running today. The core engine ## Three ways to search, one way to verify. 01 Keyword search ### Find the exact phrase, fast. Classic full-text search over every indexed document chunk, powered by Postgres's native full-text search — no separate search index to keep in sync. - Postgres full-text search over indexed chunks - Works the moment a document finishes processing - In development — targeted for the next build milestone 02 Semantic search ### Find the right idea, even with the wrong words. Documents are embedded into 768-dimension vectors and compared by meaning, not just matching text — so a question about 'remote work expenses' can surface a passage that says 'home-office reimbursement' instead. - 768-dim embeddings via Cloudflare Workers AI (BGE base) - Stored and queried with pgvector - In development 03 Hybrid search ### The best of both, blended automatically. A reciprocal-rank-fusion blend of keyword and semantic results, so an exact match and a conceptual match both surface without you having to pick a mode. - Reciprocal rank fusion (RRF) of keyword + semantic result sets - One query, one ranked list - In development 04 Question answering ### Ask a question, get an answer with receipts. Beyond search results, the planned RAG layer retrieves the most relevant chunks and generates an answer that cites exactly which document and passage it drew from — so you can verify before you rely on it. - Retrieval-augmented generation over your own documents - Citations link back to source file and passage - In development — not yet available 05 Bring your own LLM ### Use your model, not just ours, for answers. Enterprise customers will be able to point the question-answering step at their own OpenAI-compatible endpoint, with credentials stored per-organization rather than shared across tenants. - Per-organization endpoint, model name and API key - OpenAI-compatible API format - In development — enterprise tier 06 Access control ### Five roles, enforced where it actually matters. Super admin, org admin, group admin, member and viewer — and unlike a typical app-layer permission check, these are enforced by Postgres Row-Level Security policies, so the database itself won't return a row outside a user's scope. - 5-role RBAC: super\_admin, org\_admin, group\_admin, member, viewer - Enforced via Postgres RLS, not just application code - Roles and isolation are live; group-level fine-grained rules are still being built 02 Architecture ## A hosted pipeline, not a desktop app. A document you upload goes to object storage, gets picked up by an ingestion worker that splits it into overlapping chunks and embeds them, and lands in Postgres ready to be searched — by keyword, by meaning, or both. For your IT team - **Chunking** — Sliding window, ~2,048 characters per chunk with 400-character overlap - **Embeddings** — 768-dimension vectors via Cloudflare Workers AI (BGE base), stored in pgvector - **Keyword search** — Postgres full-text search over a generated tsvector column 03 Built in the open ## What's running today, and what's next. QueryLlama is pre-launch. Rather than blur the line, every capability on this site is labeled by build status — so a design partner or security reviewer knows exactly what they're evaluating today. ### Multi-tenant isolation (Postgres RLS) Every table scoped by organization; enforced at the database layer. Live in the MVP build ### Document ingestion: chunking + embeddings Upload to R2, chunk on a sliding window, embed via Cloudflare Workers AI, store in pgvector. Live in the MVP build ### Keyword search (Postgres full-text) Full-text search over indexed document chunks. In active development ### Semantic + hybrid search (RRF) Vector similarity search and a reciprocal-rank-fusion blend of keyword and semantic results. In active development ### RAG question-answering with citations Ask a question, get an answer grounded in retrieved chunks with a link back to source. In active development ### Bring-your-own-LLM (enterprise) Point the answering step at your own OpenAI-compatible endpoint. In active development ### Audit-log event writes The append-only table exists; wiring every action through to it is in progress. In active development ### Group-level permissions & tier billing Finer-grained access rules within an org, and metered plan enforcement. Designed, on the roadmap ## See it from your department's angle. Legal, HR, support and engineering all ask QueryLlama different questions. Here's what an early-access deployment looks like for each. [Explore use cases](https://royalsoftworks.com/products/queryllama/use-cases/) [How it works](https://royalsoftworks.com/products/queryllama/how-it-works/)