FAQ

FAQ & Honest Limitations

What does ragleap-rag deliberately not do (yet)?

• No agentic AI / tool-calling. This is a RAG engine — text in, grounded answer out. No agent reasoning loop, no tool-calling. That's planned as ragleap-agents, not built yet.
• No multi-step orchestration. No graph-based workflow engine (a LangGraph equivalent). Planned as ragleap-flows.
• No human-in-the-loop approval gates. A real enterprise requirement, planned as part of ragleap-agents/ragleap-flows.

These are deliberate scope boundaries, not oversights — see Roadmap for why the project stays focused rather than trying to match LangChain feature-for-feature.

Is everything in the docs actually verified?

Mostly, and where it isn't, it's labeled. 2 of 6 vector backends (pgvector, FAISS) and 3 of 8 embedding providers (Gemini, OpenAI, Ollama) are live-verified against real services. The rest are code-complete, verified against each SDK's real installed source via introspection, but not yet run against a live account — see Vector Backends.

What are the known rough edges?

• Streaming responses don't report token usage. Each provider's streaming API surfaces usage differently; doing all three correctly is separate, not-yet-done work.
• MAX_CONTEXT_CHARS is a character-count approximation (~4 chars/token for English), not an exact per-provider tokenizer count.
• Hybrid search hasn't been benchmarked for ranking-quality improvement on a corpus with genuinely conflicting dense vs. sparse rankings — only correctness (the fusion math itself, tokenization of unusual identifiers) has been verified so far.
• Transcription accuracy varies by language — the default provider (Whisper) is strongest on English/major European languages, with no built-in denoising or domain-vocabulary biasing by default.

Why does the CHANGELOG sometimes correct earlier entries instead of just updating them?

Because a changelog is a historical record, not something we quietly rewrite. When an earlier claim turns out to be wrong (it's happened — see the PineconeBackend correction in CHANGELOG.md's v0.10.1 entry), the original entry stays as-written and a new entry documents the correction. This is intentional, not an oversight.

Why is a doc/diagram fix sometimes not visible on PyPI right away?

PyPI only updates its stored README/metadata on an actual twine upload — merging to GitHub's main branch alone doesn't update the live PyPI page. If something looks fixed in the repo but not on PyPI, it's waiting on the next version publish, not broken.

Where do I report a security issue?

See SECURITY.md — please don't open a public GitHub issue for vulnerabilities.