I contribute to Haystack, deepset’s open-source agent framework. My work focuses on how agents retrieve evidence from documents and knowledge graphs.
Python
Haystack
GraphRAG
gRPC
Selected contributions#
- Graph retrieval. I built a FalkorDB retriever that uses the graph schema to translate natural-language questions into read-only Cypher queries. I also extended the document store to write graph labels and relationships from document metadata at indexing time. Agents can then query those relationships without modifying the graph.
- Embeddings. I added SPLADE sparse embedders and gRPC support for Hugging Face Text Embeddings Inference, including streaming document batches. SPLADE represents text as weighted terms for search. For models that instead produce a vector per token, I built vLLM multivector embedders for both queries and documents, preserving fine-grained matches between a question and a passage.
- Document processing. I authored the Gotenberg integration so Haystack pipelines can convert source documents to PDF through a separate service before parsing and indexing them. I extended OpenDataLoader PDF to return embedded images as Haystack documents, so figures can be retrieved alongside text. I also added asynchronous metadata extraction and concurrent embedding requests for document pipelines.