3.Summary Tree & Retriever Integration

[santo0]

PRs structure

The PRs depend on the previous ones.

1. 1.Knowledge Graph Building Pipeline  #97
2. 2.Canonicalization, Section Tree and KG Retriever #98
3. Current

SummaryEntry + summary_tree.py
Builds LLM-generated summaries bottom-up across the section tree. Leaf nodes summarize sliding windows of adjacent chunks; internal nodes summarize their children's summaries. All summaries are embedded and persisted as a FAISS index (summary_index.faiss + summary_meta.json) under the run directory.

SectionSummaryRetriever (name="section_summary")
At query time, embeds the query and searches the summary FAISS index. Each matching summary distributes its cosine similarity score to every chunk it covers; a chunk's final score is the max across all hits. Lazy-loads the embedding model on first call.

Update benchmark_retrieval.py
Evaluates all three KG retrievers (kg_node, section_tree, section_summary) plus the existing FAISS/BM25 retrievers against a query set from tests/benchmarks.yaml. Supports optional LLM relevance grading via OpenRouter.

KG and Section retrievers integration in main.py
Each of kg_node, section_tree, and section_summary is only loaded if its weight in ranker_weights is non-zero. CanonicalLookup is built once and shared across KG retrievers.

[santo0]

I'm trying to simplify this PR, I will notify when I'm done.