Living Context Graphs: Continual Semantic Learning for Enterprise Data Reasoning
Organizations: Genloop
Abstract
Enterprise AI can now perform analytics, reporting, and data-driven workflows, yet the data foundations it relies on remain hard to operationalize: business meaning, relationships, metric definitions, and reporting conventions are absent from database schemas and scattered across undocumented systems and tribal knowledge. We introduce the Living Context Graph (LCG), a continually learned context layer that makes existing enterprise data operational for AI without requiring a copy or transformation of the underlying warehouse. LCG bootstraps base semantics and relationships from the data itself, builds a concept graph that indexes business context as queries are served, writes durable semantic learnings from verified outcomes, and withdraws or re-earns context when later outcomes contradict prior routes. We evaluate LCG through text-to-SQL, a core data-reasoning task, on a transfer-aware split of three BIRD databases comprising 499 questions: 353 form the learning history and 146 are held out and never used to update the context layer. Bootstrapping alone improves overall accuracy by +4.6 percentage points. Learning only from the history raises accuracy on the untouched held-out transfer set by +22.6 points (95% CI [+14.4, +30.8]), exceeding five-shot retrieval from the same history by +12.3 points. Incremental learning reaches the one-shot endpoint within observed evaluation noise on every database. Ablating the graph changes held-out accuracy by at most one question per database but increases per-query cost 2.3-3.3x and median latency 1.8-3.4x: continually learned semantics drive accuracy, while the graph makes that context efficient to serve.
Figures & tables
| card_games | toxicology | thrombosis_pred. | pooled | ||||||||||
| Knowledge-base state | hist. | h.o. | all | hist. | h.o. | all | hist. | h.o. | all | hist. | h.o. | all | |
| 134 | 57 | 191 | 101 | 44 | 145 | 118 | 45 | 163 | 353 | 146 | 499 | ||
| No learning | |||||||||||||
| (a) | Raw schema only | 61 | 5 | 66 | 69 | 5 | 74 | 21 | 9 | 30 | 151 | 19 | 170 (34.1%) |
| (r) | 5 retrieved examples | – | 13 | – | – | 8 | – | – | 14 | – | – | 35 | – |
| (b) | Bootstrapped catalog | 71 | 6 | 77 | 79 | 4 | 83 | 23 | 10 | 33 | 173 | 20 | 193 (38.7%) |
| held-out acc. | $/query | latency (s) | ctx. recall | full hit | |
| card_games : 57 held-out, 134 learning-history questions | |||||
| 0 | 7 (12.3%) | 0.0166 | 165.1 | 87.0% | 0.0% |
| 134 | 22 (38.6%) | 0.0448 | 110.0 | 90.2% | 78.9% |
| 134 † | 23 (40.4%) | 0.1477 | 373 | 95.6% | – |
| 191 ∗ | 36 (63.2%) | 0.0402 | 182.8 | 97.0% | 91.2% |
| toxicology : 44 held-out, 101 learning-history questions | |||||
Appendix figures & tables13 assets
Supplementary material from the paper’s appendix.
Appendix
| Exp 1 | Exp 2 | Exp 3 | ||
|---|---|---|---|---|
| bootstrap | one-shot learning | online stream | ||
| C1 | bootstrapped semantic layer (§ 3.1 ) | ✓ | ✓ | ✓ |
| C2 | signal-gated graph routing (§ 3.2 ) | – | – | ✓ |
| C3 | semantic learning loop (§ 3.3 ) | – | ✓ | ✓ |
| C4a | admission gate (§ 3.4 ) | – | ✓ | ✓ |
| C4b | graph invalidation loop (§ 3.4 ) | – | – | ✓ |
| Object | What its fields carry |
|---|---|
| Column defs bootstrapped (115) | Business meaning written toward use (how the column bears on computing KPIs), a longer definition, prose on how a value is composed, sample values, and how a null must be read here. Two fields are conditional: closed-domain value lists (26/115) and a generated converter for awkward date encodings (2/115) |
| KPI defs never bootstrapped (95) | The metric expression in SQL form, so it is applied verbatim rather than paraphrased; the fully qualified columns it references; and an origin marker that reads created-with-sql on all 95, every KPI came from a query that executed, none was authored |
| User prefs never bootstrapped (83) | The rule text (a convention or a mandatory filter the SQL must respect), its scope - a real table for table-scoped rules, an org-wide marker otherwise and the concepts the rule asked to attach to |
| Concepts bootstrapped (23) | Three membership lists (columns, KPIs, preferences) that make a concept the unit of retrieval. The last is the list routing walks, so a preference is reachable only through a concept it attached to |
| on every object | A type flag - delta (a learned write, invisible at query time until admitted, kept as provenance once live) or concrete (bootstrapped, live from creation) - the upgrade-only state ladder of § 3.2 , a dedup key, parent links tying a revision to what it revises, and provenance: automated_onboarding for bootstrapped objects, the authoring question for learned ones |
| raw schema | bootstrapped | |||||
|---|---|---|---|---|---|---|
| 3 | 5 | 10 | 3 | 5 | 10 | |
| card_games (57) | 10 | 13 | 12 | 12 | 12 | 10 |
| toxicology (44) | 9 | 8 | 8 | 9 | 7 | 6 |
| thrombosis_prediction (45) | 12 | 14 | 12 | 13 | 14 | 11 |
| pooled (146) | 31 | 35 | 32 | 34 | 33 | 27 |
| with judge | without judge | |||||||
|---|---|---|---|---|---|---|---|---|
| train | h.o. | KPIs | prefs | train | h.o. | KPIs | prefs | |
| card_games | 105 | 22 | 95 | 83 | 74 | 10 | 113 | 100 |
| toxicology | 87 | 12 | 138 | 88 | 71 | 7 | 166 | 112 |
| thrombosis_prediction | 77 | 19 | 177 | 183 | 62 | 12 | 209 | 209 |
| pooled | 269 | 53 | 410 | 354 | 207 | 29 | 488 | 421 |
| card_games | toxicology | thrombosis_pred. | |
| eligible (wrong first pass) | 63 | 22 | 93 |
| fix rate | 85.7% (54) | 86.4% (19) | 80.6% (75) |
| mean attempts to fix | 3.82 | 2.90 | 4.11 |
| max attempts to fix | 6 | 5 | 8 |
| Held-out | $/query | latency (s) | ctx. recall | Graph hit | |||||
| acc. | all | full / partial | all | full / partial | all | full / partial | full | partial | |
| card_games : 57 held-out, 134 training questions | |||||||||
| 0 | 7 (12.3%) | 0.0166 | – | 165.1 | – | 87.0% | – | 0.0% | 0.0% |
| 20 | 12 (21.1%) | 0.0256 | 0.0124 / 0.0415 | 177.1 | 134 / 240 | 73.3% | 58.3 / 91.2% | 54.4% | 98.2% |
| 40 | 16 (28.1%) | 0.0288 | 0.0145 / 0.0655 | 137.2 | 122 / 252 | 69.8% | 61.6 / 90.6% | 71.9% | 98.2% |
| 60 | 13 (22.8%) | 0.0383 | 0.0170 / 0.0847 | 128.3 | 109 / 239 | 74.5% | 66.1 / 92.6% | 68.4% | 98.2% |
| break events | recovered | rate | |
|---|---|---|---|
| card_games | 19 | 10 | 53% |
| toxicology | 7 | 5 | 71% |
| thrombosis_prediction | 15 | 10 | 67% |
| pooled | 41 | 25 | 61% |
| teachers | independents | (bootstrapped) | learned (final) | |
|---|---|---|---|---|
| card_games | 70 | 64 | 29 | 46 |
| toxicology | 59 | 42 | 29 | 33 |
| thrombosis_prediction | 102 | 16 | 5 | 9 |
| database | Q | families | shared | % Q in shared fact | train / h.o. | teachers | live | pts / h.o. Q | ckpts | outcome |
|---|---|---|---|---|---|---|---|---|---|---|
| card_games | 191 | 60 | 39 | 66.5 | 134 / 57 | 70 | 39 | 1.8 | 7 | selected |
| thrombosis_prediction | 163 | 142 | 78 | 90.2 | 118 / 45 | 102 | 71 | 2.2 | 6 | selected |
| toxicology | 145 | 33 | 18 | 71.0 | 101 / 44 | 59 | 18 | 2.3 | 5 | selected |
| formula_1 | 174 | 44 | 13 | 32.8 | 132 / 42 | 15 | 13 | 2.4 | 7 | fails C4 |
| financial | 106 | 36 | 24 | 76.4 | 74 / 32 | 49 | 24 | 3.1 | 4 | fails C2, C3 |
| california_schools | 89 | 64 | 24 | 64.0 | 64 / 25 | 32 | 22 | 4.0 | 4 | fails C2, C3 |
LLM Context Management
Large language model context management studies how to select, compress, retrieve, and organize information within bounded model inputs.
See all 260 papers