AI SEARCH · 2026-11-12

AI Search — Paper Daily, 2026-11-12

Today's retrieval track mostly asks one question: how should retrieved content be handed to a model so that it actually helps? DocFS turns long documents into a file system, so an agent lists directories, searches and reads files to find evidence. OpticalRAG renders documents as images and does retrieval and compression on visual tokens. Another group of papers looks at the limits of retrieved results. CLADE finds that when a passage gives only a coarse category, models stop abstaining and start guessing. The Elastic blog post puts the permission check in a fixed step before anything reaches the LLM. On the cost side, FERN brings Text-to-SQL table selection close to LLM-based systems on some benchmarks without calling a generative LLM, and StaRSE uses static vectors that store only signs to reach very low CPU latency.

4.15

1. DocFS: File Systems as Agentic Reasoning Environments for Complex Multimodal Document Question Answering

When a VLM agent answers questions over long documents that mix text, figures and tables, the hard part is often how the document is laid out for it, not the model. DocFS takes a plain route: it turns the document into a file system, with sections as nested directories and text, figures, tables and equations stored as separate files, plus index files and relation files that record which text goes with which figure.

4.15

3. FERN: Cross-Benchmark Table Retrieval for Text-to-SQL with Frozen Retrievers and Learned Ranking

In Text-to-SQL over a catalog of hundreds of tables, the first step is finding the tables a question needs, and the popular approach is to let an LLM read the schema and pick. FERN has no generative LLM in its retrieval path: three off-the-shelf retrievers each fetch candidates into one pool, every candidate table gets a large set of features, and a LightGBM ranker trained once on BIRD's training questions orders them.

4.15

4. StaRSE: Compact Russian Sentence Embeddings with a Sign-Coded Static Encoder

When embeddings must be computed on CPUs or refreshed often, the inference cost of a Transformer encoder becomes the bottleneck. StaRSE is a static Russian sentence encoder: each token is looked up in a table to get a fixed vector, the sentence vector is the average of those vectors, and no Transformer runs at serving time. For each token, the table stores only the sign of each dimension of its direction, plus one separate length value.

3.85

5. CLADE: Category-Level Answering from Deficient Evidence in Retrieval-Augmented Language Models

Passages retrieved by RAG are often correct but coarser than the question. A question asks which specific disease a drug treats (familial hyperlipidemia), and the passage only says it treats hyperlipidemia. The authors find that when LLMs read such passages that name only the disease family, they abstain less and start guessing within the family. On biomedical questions from PrimeKG, these passages cut abstention by 13.8 points compared with passages that name neither the disease nor its family, and about half of the extra answers are wrong. Confidence filtering, a larger model and an explicit instruction to abstain all fail to stop it. Expanding the family into a candidate list with a disease ontology puts the true answer in the list 98.5% of the time once the family is right, and the real difficulty is finding that family in the passage. A passage that names the right family is not enough to answer the question.

3.80

6. Point-in-Time Encoders for Earnings Calls

In a backtest, if the retrieval encoder has read text written after the test date, future information leaks into the results. The authors train a chain of yearly encoder versions on earnings call transcripts, each reading only text dated before its cutoff year, then fine-tune them contrastively on the calls' analyst question-answer pairs. This 33.8M-parameter model retrieves answers better than every general retriever tested in every test quarter. The key point is that a version that kept training on later years retrieves better on earlier calls than the version available at the time. A control that trains on the same number of tokens without new text shows no such gain, so the gain comes from the later text. When its retrieved passages go to an LLM with the same kind of cutoff, the advantage carries over, while return forecasts barely change. Retrievers in a backtest need a time cutoff too.

3.80

7. Fine-grained authorization for Elasticsearch RAG with SpiceDB

Elastic

Enterprise knowledge-base RAG often starts with a department field as the permission filter, but it breaks once Engineering wants to share one document with Sales, or one executive needs an exception. This Elastic blog post moves permissions into a SpiceDB relationship graph: who belongs to which department and who can view which document are each stored as one relationship, and a permission check asks whether a path exists in the graph. Elasticsearch only retrieves by relevance, and before any retrieved chunk reaches the LLM, SpiceDB checks whether the current user may view it. Sharing or revoking is writing or deleting one relationship, with no reindexing. The check must sit in a fixed node the agent cannot route around, because a prompt injection could persuade the model to skip it. Filtering after retrieval leaves fewer chunks, so the authors suggest retrieving several times more candidates before filtering.

3.55

8. Graph Visualizations in Prompts for Knowledge Graph Question Answering

In knowledge graph question answering, if the retrieved triples are written as text and a drawn image of the subgraph is attached too, do VLMs answer better? The authors compare question by question on six VLMs, and adding the image raises F1 by 1.85 points on average. The gain comes mainly from precision and sits on structurally simple questions, while complex ones gain almost nothing. How the graph is drawn matters a lot. On one of the models, giving only the image without the text drops F1 by 9.42 points. A clear hierarchical layout helps, a messy layout shrinks the gain or removes it, and splitting a large subgraph into separate reasoning paths works best. Images also cost extra tokens, so the authors suggest attaching a diagram only for small, sparse subgraphs and always keeping the text triples.

3.55

9. Multimodal Abstention That Preserves Acquired Correct Answers

Adding abstention to an already fine-tuned multimodal model often turns questions it used to answer correctly into abstentions, and net accuracy hides these losses. The authors' example is CLIP-UP EmbLoRA: its net accuracy barely moves, yet underneath it loses 62 originally correct answers and fixes 41 others. ReFrain uses a linear probe on the logits of the first output token to score whether a question can be answered, then sets the threshold on a separate set of images at a point where no correct answer is lost. When the score falls below the threshold, the output is replaced with the abstain option the question already offers, and every other output stays unchanged. This way it loses no correct answer and raises the correct-abstention rate by 7.30 points. Evaluating abstention means checking which correct answers were lost, one by one, and net accuracy alone is not enough.

3.20

10. Ontology-Guided Knowledge Graph Construction and Robust Evidence Retrieval for Indian Knowledge Systems

Retrieving evidence from Indian Knowledge Systems texts is hard because one name has several transliterated spellings, names are ambiguous, and relations are spread across several passages. The authors use an LLM to extract entities and relations under an ontology and build a knowledge graph, with every relation linked back to its source passage, and retrieval walks a few bounded steps on the graph before reranking. On 320 questions from Dhara Bench, every system gets the same number of candidates. The graph alone gives the most precise top five results, and on multi-hop questions it also ranks the first relevant result highest. Merging the graph results with single-vector retrieval and other channels recovers more of the evidence. The price is speed: graph retrieval takes about 20 seconds per question on average, single-vector retrieval under 1 second. Choosing the graph or the merged results depends on whether top-rank precision or complete evidence matters more.