SLM is a full pipeline: ingest, structure, index, verify, answer, serve, and observe. Every feature exists to answer one question — can I trust what it just told me?
The pipeline
Structure-aware extraction
Hybrid BM25 + dense + FAISS
Claim-by-claim checking
Retrieval-grounded generation
OpenAI-compatible endpoint
Core engine
Every answer is generated from the retrieved passages of your corpus, then verified claim-by-claim. Unsupported statements are dropped, never invented.
Each sentence is checked against the source: entity agreement, numeric agreement, and predicate support. A wrong claim cannot survive verification.
Answers come back with section, canto/page, and the exact passage. Click any citation to open the source in context.
Lexical precision meets semantic recall, fused with reciprocal-rank fusion and a FAISS index for speed. Synonyms resolve via an entity alias map.
A pack is a folder plus a YAML manifest. The engine has zero corpus-specific code — swap packs to swap expertise with no code changes.
Anjaneya finds Hanuman. Lakshmana finds Lakshman. Ravan finds Rávan. The corpus model builds aliases, glossary, and relationships from the text itself.
Use cases
"What does our policy actually require?" — answered with the exact clause, section, and page. Audit-ready.
"Walk me through the runbook for this procedure." — grounded step-by-step in your real operating material.
"Who is X? What does the text say about Y?" — comprehensive, synthesized answers with every claim cited.
Expose any pack as an OpenAI-compatible endpoint. Point your existing agents, copilots, and workflows at it.
BYO model: local ollama, MLX on Apple silicon, transformers/PEFT on CPU, or vLLM on GPU. The engine is model-agnostic.
Toggle strict grounding to guarantee nothing is returned unless it is verified against the source. Sources-only mode returns raw passages.
Web UI
Enterprise
Self-host on a workstation, VM, or air-gapped network. Your documents never leave your control.
Login, per-pack access control, and document-level permissions. Adapters are treated as protected artifacts.
Health endpoints, request logging, eval reports, and per-pack quality grades. Monitor drift and retrain on triggers.
Adapters are versioned per pack. Roll back a bad adapter in one step; restore from daily backups of configs and indexes.
Closed-book evals on held-out chapters, graded correctness, hallucination rate, and abstention. Numbers, not adjectives.
Makefile targets for ingest, datagen, train, eval, serve, and UI. Add a pack, update a pack, roll back — all documented.