Read the source
Page-safe parsing, normalization, fingerprinting, and deterministic chunk identity.
RAG, with the reasoning left on
ATLAS/LM
Your document answers. Every claim leaves a trail.
Open workspaceINPUTPDF / TXT / MD
RETRIEVALANN + BM25 + RRF
OUTPUTANSWER + EVIDENCE + TRACE
SOURCE → CLAIM
Content-addressed
Qdrant 1.18
Strict abstention
Inspectable traces
Cost-aware modes
Scroll through the route an answer must survive before AtlasLM will show it.
Search twice
Dense ANN finds semantic neighbors. BM25 protects exact terms, names, and numbers.
Judge the evidence
RRF, reranking, MMR, and a sufficiency gate decide what deserves context.
Audit every claim
The answer is checked for known citations, coverage, grounding, and abstention.
Scroll to route the packet
LIVE PROOF / CLICK A CITATION
An answer is only the beginning.
AtlasLM combines semantic and lexical retrieval before it reranks evidence. When the source cannot support the question, strict mode stops before generation.
100% citations resolved0 unknown source labels6 selected passages
On Qdrant 1.18+, AtlasLM negotiates real 4-bit TurboQuant, then rescoring preserves retrieval quality with original vectors.
4-bit compressed indexHNSW candidate search1.5× oversampling before rescore
STRICT / ON
The most important answer is sometimes no answer.
NO EVIDENCE.
NO CLAIM.
QUERY / What was the 2028 revenue?EVIDENCE / Required fact not foundGENERATION / BLOCKED
The demo is the architecture.
Every row below is implemented, visible, and testable in the workspace.
Ingest + normalize
Content identity, metadata, deduplication
Hybrid retrieval
Dense ANN plus contextual BM25
Rank fusion
RRF, reranking, and MMR diversity
Source confidence
Strength, coverage, and agreement
Constrained generation
Source-only prompt and injection isolation
Citation audit
Claim labels checked against retrieved chunks
Hallucination fallback
Insufficient evidence blocks generation
Continuous evals
Unit suite and a live adversarial probe
Caching + memory
Document-scoped semantic answer cache
Observability
Trace ID, spans, latency, and token use
Then inspect whether the answer actually earned your trust.
Open workspace Efficient and precision modes Persistent vector index