Davide Palleschi
Back to Projects
Work
2025-04 - 2025-05

Agricultural RAG Systems

Advanced RAG pipelines using Qdrant and Qwen3 to make agricultural knowledge accessible to AI.

RAG
Vector DB
Qdrant
vLLM
NLP
Agricultural RAG Systems

Evolution: Precision Information Retrieval

Agricultural RAG Systems is the backbone of knowledge retrieval at DeepPlants. It began as Cabbage RAG, a research prototype, and matured into Cabbo RAG, a production-grade pipeline handling the indexing and retrieval of massive agronomic datasets.

The core challenge was precision. "Similarity" in vector space does not always mean "relevance" in agronomy. A paper about "wheat rust" might be embedding-similar to "corn rust," but for a farmer, the difference is total crop failure.

The Pipeline Configuration

We standardized on a high-performance, locally-hosted stack to ensure data privacy and speed.

1. Embeddings & Models

We utilized the Qwen/Qwen3-Embedding-4B model. This choice was deliberate:

  • Size vs. Performance: At 4 billion parameters, it offers a superior understanding of complex scientific syntax compared to smaller BERT-based models, while being small enough to serve efficiently.
  • Serving: The model is served via vLLM, allowing us to batch thousands of embedding requests concurrently.

2. Intelligent Chunking

Through extensive testing with USDA reports and FAOLEX legal texts, we optimized our chunking strategy:

  • Chunk Size: 500 tokens. This proved to be the "Goldilocks" zone—large enough to contain a complete semantic thought (e.g., a full regulation or a disease symptom description) but small enough to maintain precise vector focus.
  • Overlap: 64 tokens. A strategic overlap ensures that context isn't lost at the boundaries of chunks, which is critical for continuous narratives in research papers.

3. Vector Storage & Retrieval

We deployed Qdrant as our vector database. Its Rust-based engine provides exceptionally low-latency search.

  • Collection: qwen3_embedding_faolex.
  • Filtering: We heavily utilize Qdrant's payload filtering to allow agents to scope searches by document type (e.g., "Only search in USDA reports from 2024").

Handling Heterogeneous Data

The system ingest data from diverse, unstructured sources:

  • FAOLEX: Text-heavy legal documents, processed from raw .txt dumps (create_text_rag.py).
  • USDA Reports: Complex PDFs scraped and converted to Markdown (usda_md_folder), preserving header structures and tables.

The Two-Stage Reranking Advantage

To solve the "wheat rust" vs "corn rust" problem, we implemented a Two-Stage Retrieval:

  1. Dense Retrieval: Qdrant retrieves the top 100 candidate chunks based on cosine similarity.
  2. Reranking: A cross-encoder model re-scores these 100 candidates, examining the actual query-document interaction to filter out false positives. This step ensures that the top-5 chunks sent to the LLM are factually pertinent, not just textually similar.