Skip to content

Roadmap

rag-titan aims to cover every common RAG type with the same simple building blocks.

RAG types

  • Simple RAG ("simple"), available since 0.1.0
  • Adaptive RAG ("adaptive"), available since 0.1.0
  • Corrective RAG (CRAG): grade chunks, fall back to web search
  • Multi-query RAG: ask several versions of the question
  • HyDE: search with a hypothetical answer
  • Hybrid RAG: keyword (BM25) + vector search
  • Self-RAG: the model critiques its own answer
  • Parent-document RAG: search small chunks, answer with big ones
  • Graph RAG: use a knowledge graph
  • Agentic RAG: an agent decides when and what to search

Building blocks

  • LLMs: OpenAI, Anthropic, Hugging Face
  • Embedders: OpenAI, sentence-transformers
  • Vector stores: Chroma, Qdrant, FAISS, pgvector
  • Rerankers: cross-encoder, LLM reranker
  • Loaders: TXT and Markdown (TextLoader), available since 0.3.0
  • Loaders: HTML, CSV
  • Chunkers: character, token, recursive ("character", "token", "recursive"), available since 0.2.0
  • Chunkers: sentence, paragraph, heading, page ("sentence", "paragraph", "heading", "page"), available since 0.3.0
  • Chunkers: element-aware (lists, tables, code), semantic
  • Chunk offsets, page numbers and section paths, with page citations in answers, available since 0.4.0
  • Remove duplicate overlapping chunks from search results (DEDUPE_OVERLAP), available since 0.5.0
  • Evaluation helpers: retrieval and end-to-end (rag-titan evaluate), available since 0.6.0
  • Step traces and flows for every RAG type (run(), FLOW, ask --trace), available since 0.6.0
  • Swappable retrievers (BaseRetriever, DenseRetriever), available since 0.6.0
  • RAG playground: run RAG types in the browser with live flow diagrams, recorded or live with your Ollama, available since 0.7.0
  • Answer grading: deterministic answer, fact coverage, grounding and citation metrics (GenerationEvaluator), available since 0.8.0
  • Stage evaluation: chunking through end to end, scorecards, root-cause hints, reports and regression checks (RAGEvaluator, rag-titan evaluate --input), available since 0.8.0
  • Evaluation cookbook: six recipes, starter kit and animated walkthroughs, available since 0.8.1
  • Beginner evaluation helpers: golden questions template (rag-titan golden), build_dataset() for any RAG's logs, clearer summary, available since 0.8.2

Later ideas

  • LLM-based answer and citation grading
  • Stage evaluation directly from CLI runs (rag-titan evaluate --files ... --questions golden.json --stages)
  • LLM-drafted golden questions, validated by a human
  • S3 / cloud loaders
  • Confidence intervals and repeated evaluation runs
  • Scorecard breakdown by task type