AuthorsKyle Wild, Yusuke Takahashi, Asako Uraki
TL;DR
Ingest-Time Semantic Compilation (ISC) compiles a validated semantic substrate at ingest, yielding 85.2% accuracy from ~2.2k tokens vs 72.5% from ~16.3k for the best chunk baseline.
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
Query-time semantic reconstruction forces language models to re-derive corpus meaning on every query, inflating inference spend as context volume grows faster than token prices fall.
Ingest-Time Semantic Compilation shows that incremental updates can be 33.7× cheaper than full reconstruction, while compiled claims answer 85.2% of questions using ~2.2k tokens versus 72.5% using ~16.3k, exposing wasted compute, degraded accuracy, and fragile provenance in current RAG systems.
HOW IT WORKS
Ingest-Time Semantic Compilation builds a geometric layer, symbolic layer, validation gate, and index_outbox inside PostgreSQL, treating the semantic substrate as a first-class database object with its own contracts.
ISC is like adding indexes and materialized views to a database: expensive semantic work moves from read time to write time, so queries hit a compiled structure instead of scanning raw text.
This lets ISC deliver atomic, provenance-validated claims as the retrieval payload, enabling cheaper, more accurate reads than a plain context window that repeatedly re-interprets unstructured transcripts.
DIAGRAM
This diagram contrasts query-time semantic reconstruction with ingest-time semantic compilation, showing how ISC shifts work to ingest and serves compiled claims.
DIAGRAM
This diagram shows how ISC compiles transcripts, runs question answering, and compares compiled claims against chunk-based baselines and a contextualized stack.
PROCESS
Compilation Contract
ISC applies the compilation contract to each transcript, deciding which facts and embeddings to generate under a defined semantic DDL.
Validation Gate
ISC runs the validation gate, requiring exact-quote matches and correct speaker attribution before inserting rows into facts and fact_evidence.
Maintenance Contract
ISC uses index_outbox and incremental low-rank updates to maintain the geometric layer and vector index proportional to corpus change, not size.
Migration Contract
ISC applies orthogonal Procrustes alignment during embedding model upgrades, migrating the geometric layer without full re-embedding the entire corpus.
KEY CONTRIBUTIONS
Ingest-Time Semantic Compilation
ISC defines a compiled semantic substrate with a geometric layer and symbolic layer, showing incremental updates are 33.7× cheaper than full reconstruction while matching floating-point precision.
Provenance-Validated Claims
ISC’s validation gate enforces exact-quote provenance, admitting 69,746 claims and rejecting 1.1% of candidates whose quotes cannot be located in source transcripts.
Cost Model with Break-Even Reads
ISC introduces a cost model with break-even read count R*, estimating that compiling 500 documents amortizes within about 580 queries against a contextualized chunk stack.
RESULTS
QA accuracy
85.2%
+12.7 points over best chunk configuration at 72.5%
Reader tokens
2.2k tokens
vs best chunk configuration using roughly 16.3k tokens
Incremental update cost
8.4 ms per update
33.7× cheaper than 283 ms full re-decomposition
Break-even reads
≈580 queries
compilation cost equals extra query-path tokens of contextualized stack
On a held-out sample of 500 broadcast-interview transcripts and 499 questions, ISC’s compiled claims are compared against fixed-width, turn-aware, semantic chunking and a contextualized chunk stack. The 85.2% accuracy from ~2.2k tokens shows that ISC’s compiled substrate can beat chunk baselines and match a heavy contextualized stack while using about 21× fewer query-path tokens.
BENCHMARK
On a held-out sample of 500 broadcast-interview transcripts and 499 questions, ISC’s compiled claims are compared against fixed-width, turn-aware, semantic chunking and a contextualized chunk stack. The 85.2% accuracy from ~2.2k tokens shows that ISC’s compiled substrate can beat chunk baselines and match a heavy contextualized stack while using about 21× fewer query-path tokens.
BENCHMARK
Accuracy versus reader tokens per query for compiled claims, best chunk configuration, and contextualized stack.
KEY INSIGHT
ISC shows that reading more raw context can reduce accuracy: chunk reading accuracy drops from 81% to 73% as more text is supplied.
This breaks the common assumption that more retrieved tokens always help; ISC’s compiled claims maintain 98–99% evidence coverage while avoiding the degradation seen in long-context reading.
WHY IT MATTERS
ISC makes meaning a compiled, queryable substrate with integrity constraints, enabling RAG systems to treat semantics like indexed database structures instead of ephemeral context.
Builders can now design planners, maintenance pipelines, and shared semantic substrates that price compilation explicitly, validate provenance mechanically, and deliver high-accuracy answers with far fewer tokens.
Xingyu Lyu, Jianfeng He et al.
· 2026
ADAM combines Anchor extraction, Distribution estimation, Anchor selection, and Query generation to adaptively probe agent memory via an auxiliary generator and entropy based selection. On the EHRAgent benchmark with Llama2-7b-chat, ADAM reaches EQ=77 and ASR=1.00, compared to MEXTRA’s EQ=44 and ASR=0.89.
Okan Bursa
· 2026
Adaptive RAG Memory (ARM) augments a standard retriever–generator stack with a Dynamic Embedding Layer and Remembrance Engine that track usage statistics and apply selective remembrance and decay to embeddings. On a lightweight retrieval benchmark, ARM achieves NDCG@5 ≈ 0.9401 and Recall@5 = 1.000 with 22M parameters, matching larger baselines like gte-small while providing the best efficiency among ultra-efficient models.
Zhanghao Hu, Qinglin Zhu et al.
· 2026
xMemory organises agent memory via segments, memory components, groups, and dynamic maintenance with a sparsity plus semantic faithfulness objective. On LoCoMo with GPT-5 nano, xMemory reaches 38.71 BLEU and 50.00 F1 versus Nemori’s 36.65 BLEU and 48.17 F1, while cutting tokens per query from 9154.76 to 6581.20.
Answers use this explainer on Memory Papers.
Checking…