Add configurable chunking strategies for RAG document processing #1238

Activity

1. kovtcharov-amd

   added

   enhancementNew feature or requestNew feature or request

   on May 29, 2026
2. github-actions commented on May 29, 2026

   

   github-actionsBot

   on May 29, 2026 – with GitHub Actions

   Contributor

   More actions

   Good feature request, and partly in scope already — GAIA's RAG has a configurable chunking pipeline today, just not the per-document-type layer you're describing.

   What exists now (RAGConfig in src/gaia/rag/sdk.py:82):

   • chunk_size (tokens) and chunk_overlap — exactly your default.chunk_size / overlap
   • use_llm_chunking — an LLM-driven semantic strategy (_llm_based_chunking, sdk.py:907), vs. the default paragraph/sentence splitter in _split_text_into_chunks (sdk.py:1332)
   • VLM image blocks are already treated as atomic chunks (sdk.py:1512), which is the mitigation for the RAG: VLM image blocks can be split across chunks when section-based splitting path is used #786 symptom you linked

   What's genuinely missing:

   1. A named strategy enum (fixed / sentence / section / semantic) — today it's a boolean (use_llm_chunking)
   2. Section-based splitting on H1/H2 headings — no heading-aware splitter exists yet
   3. Per-document-type rules — chunk_size/strategy are global, not keyed by jira / confluence / pdf
   4. The YAML surface to express all of the above

   Suggested approach that fits existing patterns: extend RAGConfig with an optional per_type: dict[str, ChunkingConfig] and a strategy field, then refactor _split_text_into_chunks to dispatch on resolved strategy. The "document" strategy (1 chunk per item) is trivial for Jira; the "section" strategy is the only net-new splitter to write. This keeps backward compatibility since chunk_size/chunk_overlap already map cleanly onto a default config.

   One open question for a maintainer: the YAML config loader location. RAG is currently configured programmatically via RAGConfig, so where the chunking: block lives (and how openhaimer feeds it in) is an architecture call — tagging @kovtcharov-amd for that.

   Worth a contribution if you're up for it — see CONTRIBUTING.md and the RAG tuning docs at docs/sdk/sdks/rag.mdx.
3. kovtcharov-amd

   added

   ragRAG system changesRAG system changes

   domain:automationScheduler, autonomy, RAG, web search, watchers, researchScheduler, autonomy, RAG, web search, watchers, research

   on Jun 9, 2026
4. kovtcharov-amd

   added this to the vFutures milestone on Oct 19, 2026