DocuStore.io

Agentic RAG Chat

Multi-turn conversational retrieval with entity accumulation and grounding verification.

/ The problem

In traditional RAG, each question is treated independently. Ask 'What compounds target BRD4?' then follow up with 'Which of those showed selectivity?' — the system has already forgotten BRD4. Worse, without grounding verification, LLMs hallucinate plausible-sounding answers that aren't supported by the retrieved evidence.

/ How DocuStore solves it

DocuStore's conversational retrieval is an agentic loop, not a single retrieval step. The agent accumulates entities across turns, so follow-up questions inherit the compound and target context from prior exchanges. Each query is decomposed into sub-queries with targeted retrieval strategies. Retrieved passages undergo grounding verification — the agent checks that every claim in the response is supported by a cited passage before including it in the answer.

[1] ABSTRACT[2] RESULTS p.7target of CMX410?Pks13 · AT domain1is it reversible?No · Ser801 → β-lactam2covalent · irreversiblememory: CMX410 · Pks13 · AT domain✓ grounded — every claim cited

/ Pipeline

Turn 1: "What compounds target BRD4?"
  → Entity accumulation: {BRD4}
  → Search: entity-filtered retrieval for BRD4
  → Response with citations [1][2][3]

Turn 2: "Which showed selectivity over BRD2?"
  → Entity accumulation: {BRD4, BRD2, compounds from Turn 1}
  → Decompose: sub-query per compound + selectivity filter
  → Grounding check: verify each selectivity claim against passages
  → Response with citations [4][5]

Turn 3: "What was the IC50 for the most selective one?"
  → Entity accumulation: {BRD4, BRD2, top compound}
  → Structured lookup: compound-target-assay table
  → Grounding check: match IC50 value to source document
  → Cited quantitative answer

/ Key capabilities

  • Multi-turn entity accumulation preserves context across conversation turns
  • LLM-driven query decomposition for complex multi-part questions
  • Iterative retrieval loop fetches additional evidence when coverage is insufficient
  • Grounding verification ensures every claim is supported by retrieved passages
  • Inline citations link answers back to specific document pages
  • Adaptive re-retrieval when initial evidence doesn't fully answer the question
  • SSE streaming for real-time response delivery