Add graphrag-neo4j module: structured-output entity and relation extraction into a typed Neo4j graph with a vector index, multi-hop retrieval by span linking and variable-length Cypher paths, compared with plain vector RAG over the same 35 chunks and 9 questions (complete evidence 7/9 vs 5/9 at k=8); Testcontainers on neo4j:2026.10.1 and 5.26. The extraction step is a rule-based stand-in, not a language model.

Co-Authored-By: Claude Sonnet 5.5 <[email protected]> Claude-Session: https://claude.ai/code/session_01CifR2BE7PBvRy5hT8TS9oz
2026-10-26 02:03:40 +00:00
parent 279be7eb9f
commit 108fc44542
66 changed files with 1997 additions and 0 deletions
@@ -29,5 +29,6 @@ Runnable companion code for the Spring AI articles on [ankurm.com](https://ankur
 | [`hybrid-search/`](hybrid-search) | Vector, `tsvector` and a Java Okapi BM25 over one PostgreSQL table (pgvector HNSW + GIN), fused with Reciprocal Rank Fusion in Java and in a single SQL statement, behind a `DocumentRetriever` for `RetrievalAugmentationAdvisor`. 60 questions in four kinds over 94 passages: hybrid helps exact error codes and hurts paraphrases; `k` tuned on the evaluation set did not hold up on held-out folds; plans and latency at 94 and 50,094 rows. Local all-MiniLM-L6-v2. Spring Boot 4.1.2, Spring AI 2.0.2, Java 25. | [Hybrid Search in Java: BM25 + Vector with PostgreSQL and Reciprocal Rank Fusion](https://ankurm.com/hybrid-search-java-pgvector-tsvector-reciprocal-rank-fusion/) |
 | [`rerank/`](rerank) | A bi-encoder (all-MiniLM-L6-v2) retrieves candidates, a cross-encoder (ms-marco-MiniLM-L-6-v2) rescores them, and a `DocumentPostProcessor` keeps the top 5 for `RetrievalAugmentationAdvisor`. Both models run in-process through ONNX Runtime 1.31.0 with DJL tokenizers 0.39.0 (no database, no API key). 60 questions over 94 passages: hit@1 41 to 49 and MRR 0.78 to 0.87, almost all of it on exact error codes, with paraphrases slightly worse; pool size, batch size, threads and int8 against latency on 2 vCPUs; and five traps reproduced (missing `token_type_ids`, swapped pair order, logits are not probabilities, truncation, a pool that is too small). Spring Boot 4.1.2, Spring AI 2.0.2, Java 25. | [Reranking RAG Results in Java: Cross-Encoder Reranker with ONNX Runtime Before the LLM](https://ankurm.com/reranking-rag-results-java-cross-encoder-onnx-runtime-spring-ai/) |
 | [`audio/`](audio) | A meeting-minutes generator: `OpenAiAudioTranscriptionModel` -> `ChatClient.entity(Minutes.class)` -> `OpenAiAudioSpeechModel`. Splitting a 30 MB recording at a silence under the 25 MB upload limit (and what a blind split does), SDK retries that re-upload the file, the 4096-character speech limit, three replies that fail to parse, and a cost model with formulas. The audio endpoints are a local mock and the chat model a stub; no real model was called. Spring Boot 4.1.2, Spring AI 2.0.2, Java 25. | [Speech-to-Text and Text-to-Speech with Spring AI: Build a Meeting-Minutes Generator](https://ankurm.com/spring-ai-speech-to-text-text-to-speech-meeting-minutes/) |
+| [`graphrag-neo4j/`](graphrag-neo4j) | Entities and relations extracted with `ChatClient.entity(...)` into a typed Neo4j graph with a vector index on entity names (Testcontainers, `neo4j:2026.10.1-community`; also run on 5.26), then multi-hop retrieval by span linking, variable-length Cypher paths and path ranking, against plain vector RAG over the same 35 chunks and 9 questions: complete evidence 7/9 for the graph, 5/9 for vector k=8, 2/9 for k=4, with the traps reproduced (duplicate entity under an alias, an embedding merge that joins the wrong incidents, a reply with a sentence before the JSON fence, a hostile relation type, hop explosion). **The extraction step is a rule-based stand-in, not a language model**; Neo4j, embeddings and Cypher are real. Spring Boot 4.1.2, Spring AI 2.0.2, Java 25. | [GraphRAG in Java with Neo4j and Spring AI](https://ankurm.com/graphrag-java-neo4j-spring-ai-knowledge-graph/) |

 Upgrading from Spring AI 1.x: [migration guide](https://ankurm.com/spring-ai-1-to-2-migration-guide/).