LLM-generated community summaries for LightRAG high-level retrieval

[ronaldmannak]

Motivation

LightRAG's high-level (community) store is only as good as its summaries. Today LightRAG.communitySummary produces a thin, deterministic string — member entity names plus the relationship types connecting them. The real GraphRAG/LightRAG "global" technique summarizes each community with an LLM, which makes theme/global questions much stronger.

Scope

• Add an LLM-backed community summarizer that, given a Community + KnowledgeGraph, produces a natural-language summary of the members, their relationships, and evidence.
• Add the prompt to Prompts (mirroring the existing prompt style).
• Keep the current entity/relationship string as the deterministic offline fallback when no LLM is available (so offline behavior and tests stay deterministic).
• Wire it into LightRAG.communitySearcher (the community documents are what gets embedded/searched).

Files (parallel-safe)

Sources/GraphRAG/LightRAG/Searchers.swift (+ maybe a new CommunitySummarizer.swift), Sources/GraphRAG/Entity/Prompts.swift. Minimal overlap with other issues; if it needs a config flag, coordinate with the "Builder ergonomics" issue.

Acceptance criteria

• With an LLM available, community documents contain LLM-generated summaries; without one, the existing deterministic summary is used unchanged.
• New unit test for the offline fallback path; the end-to-end offline test still passes.
• CI green.

Context: follow-up to LightRAG (PR #2) / config inheritance (PR #3).