Open source · Self-hosted · Bring your own LLM

Chemistry-aware document intelligence

DocuStore extracts knowledge from your documents and presentations. Search with chemical awareness that understands biochemistry and molecular structures. Ask questions and get answers cited to the source.

Get started View on GitHub

01
Extract
OCSR reads molecular structures into SMILES; hybrid NER pulls compounds, targets, and assays out of prose.
02
Index
Dense, sparse, and structural vectors — three representations so identifiers and chemistry both stay findable.
03
Search
Hard entity filtering, RRF fusion, and cross-encoder reranking cut context pollution before it reaches you.
04
Chat
Multi-turn entity memory, query decomposition, and grounding checks — every answer carries its citations.

Intelligence layers, extract to answer

3×

Vector reps per document

RAG failure modes eliminated

/ How it works

One pipeline, from raw documents to cited answers

// image → structure

SMILES: CC(=O)Oc1ccccc1C(=O)O

confidence 0.97

Domain extraction

OCSR for molecular structures, hybrid NER for entities and relationships, multi-level summaries.

Multi-rep indexing

Dense plus sparse text vectors, SMILES embeddings via ChemBERTa, summary vectors for hierarchy.

Learn more

BMS-986158filter

BRD423 → 4

Entity-aware search

Hard entity filtering, hybrid RRF fusion, cross-encoder reranking, structured bioactivity lookup.

Grounded chat

Multi-turn entity memory and grounding checks — every answer comes back with citations.

Learn more

Works with cloud LLMs like OpenAI, Gemini, and Claude. Run local models via Ollama.

LLM setup

/ Why generic RAG fails

Five failure modes, fixed at the root

01
Vocabulary gap
Opaque compound IDs like SACC-3060 are meaningless to embeddings. Sparse indexing with identifier-preserving tokenization bridges the gap.
02
Context pollution
Unrelated compound-target mentions contaminate retrieval. Hard entity filtering before retrieval eliminates the noise.
03
Conversational drift
Each query treated independently loses entity context across turns. Multi-turn entity accumulation keeps the thread.
04
Structure opacity
Chemical structures are images, invisible to text search. OCSR extracts SMILES and indexes them separately.
05
Data blindness
Quantitative bioactivity relationships like IC50 and selectivity are lost in prose. Deterministic lookup of compound-target-assay records recovers them.