DocuStore.io
by sidx

Introducing DocuStore

An open-source document intelligence platform built for drug discovery research.

We're excited to announce DocuStore — an open-source document intelligence platform purpose-built for drug discovery research.

The Problem

Pharmaceutical R&D generates enormous volumes of internal documents: experimental reports, patent analyses, presentations, and study summaries. These documents contain critical institutional knowledge — compound structures, bioactivity data, target-compound relationships — that's effectively invisible to traditional search.

Generic RAG systems fail on this domain for five specific reasons:

  1. Vocabulary gap: Compound identifiers like SACC-3060 or BMS-986158 are meaningless to general-purpose embeddings.
  2. Context pollution: Documents mention dozens of compounds, and generic retrieval can't distinguish which mentions are relevant.
  3. Conversational drift: Follow-up questions lose entity context from prior turns.
  4. Structure opacity: Chemical structures are images, not text — they can't be searched.
  5. Data blindness: Quantitative bioactivity relationships (IC50 values, selectivity data) are trapped in prose.

Our Approach

DocuStore addresses these failures with a four-layer architecture:

Layer 1 — Domain-Aware Extraction: OCSR (Optical Chemical Structure Recognition) extracts molecular structures from page images and converts them to SMILES. Hybrid NER (dictionary + LLM) extracts entities and relationships.

Layer 2 — Multi-Representation Indexing: Three Qdrant vector collections capture different facets: dense+sparse text embeddings, ChemBERTa SMILES embeddings, and summary embeddings for hierarchical retrieval.

Layer 3 — Entity-Constrained Retrieval: Hard entity filtering before similarity search eliminates context pollution. Hybrid RRF fusion and cross-encoder reranking ensure precision.

Layer 4 — Agentic Conversational RAG: Multi-turn entity accumulation, query decomposition, and grounding verification produce cited answers grounded in source documents.

Built on Proven Infrastructure

DocuStore is event-sourced (EventStoreDB), uses CQRS with MongoDB read models, and orchestrates all processing workflows through Temporal. This gives us:

  • Complete audit trail of every document processing step
  • Durable, retryable workflows that survive crashes
  • Clean separation between write and read paths
  • Observable pipeline with Temporal dashboard

Open Source

DocuStore is open source under the MIT license. Deploy it with Docker Compose in minutes.

We'd love your feedback. Star the repo, file issues, and let us know what you'd like to see next.