Skip to content

Instantly share code, notes, and snippets.

View crajah's full-sized avatar

Chandan Rajah crajah

3 followers · 7 following
View GitHub Profile
@crajah
crajah / results-at-a-glance.csv
Last active August 25, 2026 06:58
post-graph-rag: results at a glance — LongMemEval, LightRAG, ECT-QA
overall multi-session temporal knowledge-update
post-graph-rag · gemini-3.7-flash 68.3% 66.2% 52.6% 70.5%
Zep/Graphiti · gpt-4o 71.2% 57.9% 62.4% 83.3%
Zep/Graphiti · gpt-4o-mini 63.8% 40.6% 36.5% 76.9%
Full-context baseline · gpt-4o 60.2% 44.3% 45.1% 78.2%
@crajah
crajah / ectqa-scoring-and-prompt-changes.csv
Last active August 23, 2026 23:09
post-graph-rag on ECT-QA — accuracy by scoring metric and prompt change
accuracy mean numeric F1 refused
judge panel, original prompt 0.162 — 75%
numeric F1, same answers 0.221 0.191 75%
+ extraction split, top_k 48, partial answers 0.324 0.447 7%
+ scope discipline 0.353 0.451 10%
@crajah
crajah / longmemeval-paired-ablation.csv
Created August 23, 2026 12:25
post-graph-rag on LongMemEval — paired ablation of MMR and node-distance reranking
accuracy delta
baseline 60% —
MMR 65% +5
node distance 70% +10
@crajah
crajah / longmemeval-by-question-type.csv
Last active August 23, 2026 18:33
post-graph-rag vs Zep on the full LongMemEval (oracle, 499/500) — accuracy by question type
question type n post-graph-rag (flash) Zep (gpt-4o) Zep (gpt-4o-mini)
single-session-user 70 91.4% 92.9% 81.4%
single-session-assistant 55 89.1% 80.4% 81.8%
knowledge-update 78 70.5% 83.3% 76.9%
multi-session 133 66.2% 57.9% 40.6%
temporal-reasoning 133 52.6% 62.4% 36.5%
single-session-preference 30 50.0% 56.7% 30.0%
overall 499 68.3% 71.2% 63.8%
@crajah
crajah / pg19-property-graphs.html
Created August 14, 2026 11:39
PostgreSQL 19 Property Graphs × post-graph — architecture exploration
---
layout: null
---
<!DOCTYPE html>
<html lang="en">
<head>
<meta charset="utf-8">
<meta name="viewport" content="width=device-width, initial-scale=1">
<title>PG19 Property Graphs × post-graph</title>
<style>
@crajah
crajah / retrieval-quota-by-question-shape.csv
Created August 13, 2026 08:54
post-graph-rag: blind A/B wins for relation_seed_quota 0.5 vs 1.0, by question shape
Question shape quota 0.5 quota 1.0
Entity — what role did the FAA play 8 4
Thematic — what caused cash flow to turn negative 7 4
Chain — how did regulatory action translate into financial consequences 2 6
@crajah
crajah / relation-channel-on-topic-share.csv
Last active August 13, 2026 09:55
post-graph-rag: on-topic share of retrieved relations, traversal only vs adding relation-embedding search
On-topic share of relations reaching the prompt quota 0.0 0.5 1.0
gpt-oss-120b 49% 69% 73%
MiniMax-M2.7 66% 88% 98%
@crajah
crajah / capability-comparison.csv
Last active August 23, 2026 12:25
Neo4j Apache AGE post-graph
Runs inside your PostgreSQL No Yes Yes
openCypher Yes Yes Partial — documented subset
Tables you may index and constrain n/a No — discouraged Yes
Vector search on vertices and edges via plugin No Yes
Shadow audit log of every mutation enterprise No Yes
Append-only history per element No No Yes
Schema-per-tenant isolation enterprise per graph Yes
Needs a recent PostgreSQL n/a Partial — 11–16 Yes — any supported
@crajah
crajah / retrieval-recall-granularity-fix.csv
Created August 11, 2026 21:54
Question before after
Deferred production costs arc 9 8
What turned cash flow negative 2 13
737 ↔ cash flow over time 5 44
FAA role across filings 19 31
@crajah
crajah / multi-hop-retrieval-by-depth.csv
Created August 11, 2026 21:54
Question 1 hop 2 hops 3 hops
Deferred production costs arc 7 (100% on-topic) 30 (47%) 202 (10%)
What turned cash flow negative 13 (77%) 20 (60%) 59 (39%)
737 ↔ cash flow over time 32 (91%) 125 (48%) 435 (23%)
FAA role across filings 31 (74%) 73 (52%) 307 (18%)