Skip to content

Graph retrieval — research

Thirteen investigations — twelve external projects plus the in-house precedent — and the synthesis that turns them into a gated build order for the links release and the graph release. Research, not specification: where these disagree with ../DESIGN.md, DESIGN wins; where they disagree with ../../plans/20260729_0256-links-and-graph.md about what is built when, the plan wins.

Built, and its channel is off. The links release this research shaped shipped in 0.5.0 and 0.6.0, and the graph release completed in 0.11.0 — G1 and G4 in 0.6.0, G2 in 0.7.0, G3, G5 and G6 in 0.11.0. G5's gate ran and did not pass, so graph_channel ships off: the structure is built and measured, and nothing in the staged channels below is licensed by that result. ../STATUS.md is the authority on what exists.

Start here

Doc What it is
PINAKES_APPROACH.md The synthesis. The Pinakes graph: lazy, agent-driven, budget-tunable. §9 is the eval gate each channel must pass; §10 is its build order, sequenced into increments by plans/20260729_0256-links-and-graph.md. Five adversarial passes
GRAPH_RAG.md The research record — what the literature actually says about graphs, RAG and agents, and what Pinakes should take from it

Investigations

Licences are recorded because they decide whether an idea may become code here — see the gate below.

Project Licence Why it was read
ClaudeKB in-house What was already tried here, and what it cost
LightRAG MIT Dual-level retrieval — the most-copied simple design
Microsoft GraphRAG MIT The reference implementation, and its own retreat from eager indexing into LazyGraphRAG
Graphiti Apache-2.0 Temporal knowledge graphs, bitemporal edges
HippoRAG 2 MIT Personalised PageRank over a passage graph — the PPR channel's origin
fast-graphrag MIT The cost profile of doing PPR cheaply
Graph-R1 MIT Agentic multi-turn traversal — closest to the agent-driven posture
datastax/graph-rag Apache-2.0 Retrofitting edges onto an existing vector store
code-graph-rag MIT Structural edges that need no LLM
MiniRAG MIT Whether a small local model suffices for heterogeneous graph indexing
LinearRAG GPL-3.0 Linear-cost construction — whether the build step can stay free
LogicRAG GPL-3.0 Reasoning over logical structure — the upper bound on query complexity
Youtu-GraphRAG academic only Schema-guided hierarchical construction — schema as a cost control

The licence gate

These are read-only references. Never copy code from the three marked ⛔.

  • LinearRAG and LogicRAG are GPL-3.0.
  • Youtu-GraphRAG ships a custom licence forbidding commercial and production use. Its README badge says MIT and is wrong — the LICENSE file is the truth.

Pinakes is Apache-2.0 and pnk serve is a network service, so a copyleft obligation would reach anyone embedding it. The same reasoning ruled out PyMuPDF in favour of pypdfium2 (plans/20260727_1543-v0.2.md, decision 1). Ideas and measured results are free to reuse; source is not — read them, then write it yourself. Stated once in PINAKES_APPROACH.md §"License gate" and repeated here because the index is where someone starts.