ref. citation-grounded knowledge engine

Every answer, bound to its source.[1]

A knowledge engine with a self-correcting retrieval pipeline — it grades its own context, fixes weak queries, and verifies every answer is grounded before you see it.

[1] every claim links back to where it came from.
the loop / retrieve grade correct verify cite
fig.01 — retrieval-graph grounded
sources engine answer correct & re-retrieve doc · 01 0.71 doc · 02 0.94 doc · 03 0.62 RAG grade·verify response [1] [2] grounded · 2 cited
sources · 3 graded verified · 0 ungrounded
/built on a production stack
LangGraph LangGraph
/built like production, not a demo

Three properties that separate a real engine from a clever prototype.

fig.01

Self-correcting retrieval

Grades its own context, rewrites weak queries, falls back to web search, and loops — retrieve → grade → retry — until it's confident the evidence is good enough to answer.

relevance grading query rewrite web fallback
fig.02

Grounded & cited

Every claim traces to a source. A verification step catches hallucination before the answer ever ships.

fig.03

Quality you can measure

A built-in evaluation harness scores faithfulness and retrieval precision, gated in CI so quality can't silently regress.

faithfulness across 8 CI runs
/the engine

A corrective-RAG loop, not a single pass.

Weak context is detected and repaired before generation. Ungrounded drafts are sent back. The answer only leaves the pipeline once it's verified.

rewrite query · re-retrieve not grounded → regenerate
web search
when context is weak
/01
retrieve
vector + keyword
/02
grade documents
relevance check
/03
generate
with graded context
/04
verify grounding
anti-hallucination
/05
cited answer
sources attached
/proven by evaluation, not vibes

Answer quality is measured, tracked, and gated in CI.

Faithfulness
0.94
Context Precision
0.89
Context Recall
0.87
Answer Relevancy
0.92
// naive RAG vs corrective RAG
Naive
Corrective
Δ
Faithfulness
0.71
0.94
+0.23
Context Precision
0.68
0.89
+0.21
Hallucination rate
0.14
0.03
−0.11
Answer Relevancy
0.79
0.92
+0.13
placeholder values — gated against 180 eval questions on every CI run.
/the numbers
320ms
p50 query latency
0.94
faithfulness score
18%
self-correction trigger rate
12.4k
sources indexed in the demo
180
eval questions gating CI
/ try it

See it answer your questions — with receipts.