Query-Aware Flow Diffusion for Graph-Based RAG with Retrieval Guarantees

Baruch Gutow; Davoud Ataee Tarzanagh; Hankyu Moon; Heng Hao; Masoud Faraki; Oxana Verkholyak; Seungjai Min; Sima Didari; Wenjun Hu; Zhuoping Zhou

arxiv: 2605.18775 · v1 · pith:SDO6T7TMnew · submitted 2026-04-21 · 💻 cs.IR · cs.AI

Query-Aware Flow Diffusion for Graph-Based RAG with Retrieval Guarantees

Zhuoping Zhou , Davoud Ataee Tarzanagh , Sima Didari , Wenjun Hu , Baruch Gutow , Oxana Verkholyak , Masoud Faraki , Heng Hao

show 2 more authors

Hankyu Moon Seungjai Min

This is my paper

Pith reviewed 2026-05-21 01:08 UTC · model grok-4.3

classification 💻 cs.IR cs.AI

keywords graph-based RAGquery-aware flow diffusionretrieval guaranteessubgraph recoveryquery embedding alignmentmulti-hop reasoningtraining-free framework

0 comments

The pith

Query-aware flow diffusion recovers relevant subgraphs with high probability under mild signal-to-noise conditions.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper introduces Query-Aware Flow Diffusion RAG (QAFD-RAG), a training-free approach that modifies graph traversal so edges are weighted by how closely their endpoints match the query embedding. This guides exploration toward semantically relevant paths instead of relying on fixed neighborhoods or communities. A sympathetic reader would care because existing graph-based RAG methods often lack guarantees on the quality of retrieved subgraphs, limiting reliability for multi-hop reasoning tasks. The method supplies the first statistical assurances that relevant subgraphs can be recovered with high probability, while keeping computation proportional to the size of the useful subgraph rather than the entire graph.

Core claim

QAFD-RAG performs query-aware traversal in which edges receive dynamic weights based on the alignment between their endpoint embeddings and the query embedding. Under mild signal-to-noise conditions on the graph, this process recovers the relevant reasoning subgraph with high probability. The algorithm converges exponentially fast and its complexity scales with the retrieved subgraph size rather than the full graph size.

What carries the argument

Query-aware flow diffusion, which dynamically reweights edges during traversal according to their semantic alignment with the query embedding to steer flow toward relevant regions.

If this is right

The framework supplies the first statistical guarantees for query-aware graph retrieval rather than heuristic subgraph selection.
Retrieval complexity grows with the size of the useful subgraph instead of the full knowledge graph.
The same query-aware weighting can be applied to question-answering and text-to-SQL tasks to produce more relevant reasoning chains.
Exponential convergence holds whenever the mild signal-to-noise condition is satisfied.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

If the alignment-based weighting works, similar dynamic reweighting could be tested on graphs with time-varying edges to handle evolving knowledge.
The separation of semantic flow from pure structural connectivity suggests the method may reduce retrieval of spurious multi-hop paths in densely connected but loosely related domains.
Extending the guarantees beyond the current mild conditions would require characterizing how embedding quality affects the required signal strength.

Load-bearing premise

Mild signal-to-noise conditions exist such that query embedding alignment with relevant edges is strong enough to guide diffusion away from structurally connected but semantically irrelevant parts of the graph.

What would settle it

Running the algorithm on a graph where query embeddings show only weak alignment with the true relevant edges, even under low overall noise, and observing that the probability of recovering the relevant subgraph drops below the claimed high-probability threshold.

Figures

Figures reproduced from arXiv: 2605.18775 by Baruch Gutow, Davoud Ataee Tarzanagh, Hankyu Moon, Heng Hao, Masoud Faraki, Oxana Verkholyak, Seungjai Min, Sima Didari, Wenjun Hu, Zhuoping Zhou.

**Figure 2.** Figure 2: Two-stage QAFD-RAG framework: the indexing stage builds a KG from documents, and [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** Figure 3: Sensitivity of QAFD-RAG to the number of seed nodes (left), initial mass [PITH_FULL_IMAGE:figures/full_fig_p022_3.png] view at source ↗

**Figure 4.** Figure 4: Left: Schema path identified by QAFD-RAG for SQL generation on query local039 (Lei et al., 2024). Right: Stepwise schema exploration for the same query using Spider-Agent (ReAct) (Lei et al., 2024). QAFD-RAG discovers the full reasoning path in one pass, sharply reducing LLM calls. In contrast, Spider-Agent incrementally explores the schema, requiring 14 calls and higher latency. This comparison highlights… view at source ↗

**Figure 5.** Figure 5: QAFD-RAG retrieved schema for Query-Local141 in the AdventureWorks database KG [PITH_FULL_IMAGE:figures/full_fig_p024_5.png] view at source ↗

read the original abstract

Graph-based Retrieval-Augmented Generation (RAG) systems leverage interconnected knowledge structures to capture complex relationships that flat retrieval struggles with, enabling multi-hop reasoning. Yet most existing graph-based methods suffer from (i) heuristic designs lacking theoretical guarantees for subgraph quality or relevance and/or (ii) the use of static exploration strategies that ignore the query's holistic meaning, retrieving neighborhoods or communities regardless of intent. We propose Query-Aware Flow Diffusion RAG (QAFD-RAG), a training-free framework that dynamically adapts graph traversal to each query's holistic semantics. The central innovation is query-aware traversal: during graph exploration, edges are dynamically weighted by how well their endpoints align with the query's embedding, guiding flow along semantically relevant paths while avoiding structurally connected but irrelevant regions. These query-specific reasoning subgraphs enable the first statistical guarantees for query-aware graph retrieval, showing that QAFD-RAG recovers relevant subgraphs with high probability under mild signal-to-noise conditions. The algorithm converges exponentially fast, with complexity scaling with the retrieved subgraph size rather than the full graph. Experiments on question answering and text-to-SQL tasks demonstrate consistent improvements over state-of-the-art graph-based RAG methods.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper adds query-aware dynamic edge weighting to graph RAG and claims the first statistical recovery guarantees with exponential convergence, but those guarantees rest on an unverified mild signal-to-noise assumption that may not prevent flow leakage in practice.

read the letter

The main takeaway is that this work introduces Query-Aware Flow Diffusion RAG, which dynamically weights edges during traversal according to how well their endpoints align with the query embedding. It claims this produces relevant subgraphs with high probability under mild conditions, exponential convergence, and complexity that scales only with the retrieved subgraph rather than the full graph. That combination is positioned as an advance over static neighborhood or community exploration in prior graph RAG methods. If the guarantees hold, the approach could give multi-hop retrieval a bit more theoretical footing than the usual heuristics. The training-free design and reported gains on question answering and text-to-SQL tasks are straightforward positives that show the method is at least implementable and competitive in the tested settings. The framing of the problem, highlighting how existing graph methods ignore query semantics, is clear and on target. The soft spot is the load-bearing assumption that query embedding alignment will reliably steer flow away from structurally connected but irrelevant regions. The stress-test note is right to flag this: without a concrete SNR threshold or robustness checks against partial similarities on noise edges, the high-probability recovery and subgraph-size scaling claims could break down in realistic graphs where embeddings are noisy. The abstract supplies no equations, derivation steps, or proof sketch, so it is impossible to tell whether the guarantees are derived cleanly or depend on self-referential conditions. This paper is aimed at information retrieval researchers working on graph-based RAG who want something beyond pure heuristics. A reader interested in theoretical support for multi-hop retrieval would get value from the query-aware idea even if the math needs tightening. It has enough structure and engagement with the literature to deserve a serious referee who can examine the full derivations, experimental protocols, and assumption checks. I would send it to peer review.

Referee Report

2 major / 1 minor

Summary. The paper proposes Query-Aware Flow Diffusion RAG (QAFD-RAG), a training-free graph-based RAG framework that dynamically weights edges during traversal according to their alignment with the query embedding. It claims to deliver the first statistical guarantees for query-aware subgraph retrieval, recovering relevant subgraphs with high probability under mild signal-to-noise conditions, with exponential convergence and per-query complexity scaling only with the size of the retrieved subgraph rather than the full graph. Experiments on question-answering and text-to-SQL tasks report consistent gains over prior graph-based RAG baselines.

Significance. If the claimed recovery guarantees and complexity bounds can be rigorously established, the work would meaningfully advance graph-based RAG by replacing heuristic traversal with a query-adaptive diffusion process that carries explicit probabilistic recovery statements. The training-free design and subgraph-size scaling are practically attractive for large knowledge graphs. The significance is tempered by the absence of an explicit SNR threshold or robustness margin in the provided description.

major comments (2)

[Theoretical Guarantees] The central high-probability recovery claim rests on the unquantified assumption that query-embedding alignment is sufficient to confine diffusion to relevant components and prevent leakage into structurally connected but semantically irrelevant regions. No explicit SNR threshold, misalignment tolerance, or leakage-probability bound is supplied, which directly undermines the exponential-convergence and subgraph-size complexity statements.
[Theoretical Analysis] The manuscript states that the algorithm 'converges exponentially fast' and that complexity 'scales with the retrieved subgraph size,' yet supplies neither a proof sketch nor the precise definition of the signal-to-noise ratio under which these statements hold. Without these derivations the statistical guarantees cannot be evaluated.

minor comments (1)

[Abstract] The abstract introduces 'mild signal-to-noise conditions' without a forward reference to the theorem or definition that makes the term precise.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the careful reading and constructive comments on our work. We address each major comment below with clarifications drawn directly from the manuscript's analysis and commit to revisions that strengthen the explicitness of the theoretical statements without altering the core claims.

read point-by-point responses

Referee: [Theoretical Guarantees] The central high-probability recovery claim rests on the unquantified assumption that query-embedding alignment is sufficient to confine diffusion to relevant components and prevent leakage into structurally connected but semantically irrelevant regions. No explicit SNR threshold, misalignment tolerance, or leakage-probability bound is supplied, which directly undermines the exponential-convergence and subgraph-size complexity statements.

Authors: The manuscript defines the relevant signal-to-noise conditions via the expected alignment gap between query-relevant and irrelevant edges, which is used to bound the probability that flow leaks outside the target subgraph. We agree that an explicit numerical threshold and a derived leakage bound would improve clarity. In the revision we will add a formal SNR definition (as the ratio of mean alignment difference to its variance) together with a high-probability leakage bound obtained from a standard Chernoff-type argument on the diffusion process. This will also render the exponential convergence rate and subgraph-size complexity explicit. revision: yes
Referee: [Theoretical Analysis] The manuscript states that the algorithm 'converges exponentially fast' and that complexity 'scales with the retrieved subgraph size,' yet supplies neither a proof sketch nor the precise definition of the signal-to-noise ratio under which these statements hold. Without these derivations the statistical guarantees cannot be evaluated.

Authors: The exponential convergence is a consequence of the contraction mapping property of the query-aware diffusion operator once the alignment scores satisfy the mild SNR condition stated in the paper; the complexity claim follows because the traversal halts once the relevant component is recovered. We acknowledge that the current version presents these statements at a high level without a self-contained proof sketch or an expanded SNR definition. We will insert a concise proof outline (using martingale concentration on the cumulative flow) and the precise SNR expression into the revised appendix. revision: yes

Circularity Check

0 steps flagged

No circularity detected in derivation chain

full rationale

The abstract and available description present QAFD-RAG as a training-free framework whose statistical recovery guarantees are stated to hold under external mild signal-to-noise conditions on query embedding alignment. No equations, fitted parameters, self-citations, or ansatzes are exhibited that would reduce the claimed exponential convergence or subgraph-size complexity back to the inputs by construction. The central claims therefore remain independent of any self-referential reduction and rest on the stated assumptions rather than on renaming or re-deriving the same quantities.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the domain assumption of mild signal-to-noise conditions that make query embedding alignment effective for guiding relevant flow; no free parameters or invented entities are visible in the abstract.

axioms (1)

domain assumption Mild signal-to-noise conditions in the graph data allow query embeddings to distinguish relevant paths from irrelevant but connected regions.
Invoked to support the high-probability recovery guarantee for query-aware subgraphs.

pith-pipeline@v0.9.0 · 5774 in / 1327 out tokens · 37344 ms · 2026-05-21T01:08:26.651266+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

query-aware edge weights ¯w(q,u,v) := c·Hsim(h(u),h(v)) ◦ (a + b·(Hsim(h(u),h(q)) ◦ Hsim(h(v),h(q)))) (Eq. 1); recovery Rk ⊆ supp(x*) with bounded leakage under ˆµ/ˆσ = ω(√(d log |V|)) (Thm 7)
IndisputableMonolith/Foundation/ArithmeticFromLogic.lean embed_strictMono_of_one_lt unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

exponential convergence O(¯d · ∥∆∥1 · log(1/ϵ)) scaling with subgraph size (Cor. 4)

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

32 extracted references · 32 canonical work pages

[1]

doi: 10.18653/v1/2024.findings-emnlp.685

Association for Computational Linguistics. doi: 10.18653/v1/2024.findings-emnlp.685. URLhttps://aclanthology.org/2024.findings-emnlp.685/. Parth Sarthi, Salman Abdullah, Aditi Tuli, Shubh Khanna, Anna Goldie, and Christopher D Manning. Raptor: Recursive abstractive processing for tree-organized retrieval. InThe Twelfth International Conference on Learning...

work page doi:10.18653/v1/2024.findings-emnlp.685 2024
[2]

Saba Sturua, Isabelle Mohr, Mohammad Kalim Akram, Michael Günther, Bo Wang, Markus Krimmel, Feng Wang, Georgios Mastrapas, Andreas Koukounas, Nan Wang, et al

doi: 10.1137/080744888. Saba Sturua, Isabelle Mohr, Mohammad Kalim Akram, Michael Günther, Bo Wang, Markus Krimmel, Feng Wang, Georgios Mastrapas, Andreas Koukounas, Nan Wang, et al. jina-embeddings-v3: Multilingual embeddings with task lora.arXiv preprint arXiv:2409.10173, 2024. Heli Sun, Fang He, Jianbin Huang, Yizhou Sun, Yang Li, Chenyu Wang, Liang He...

work page doi:10.1137/080744888 2024
[3]

uses multi-stage retrieval and ego-network aggregation. PathRAG (Chen et al., 2025a) further improves graph-based RAG by retrieving key relational paths rather than redundant neighborhood information, using flow-based pruning to identify reliable paths and strategic prompt organization to enhance LLM responses. However, all aforementioned methods still st...

work page 2013
[4]

All other hyperparameters remain fixed across experiments

(4096-dim), optimized for retrieval tasks; and (iv) GritLM-7B (Muennighoff et al., 2024) (4096-dim), a unified generative-embedding model. All other hyperparameters remain fixed across experiments. Table 7 presents results across five evaluation dimensions. QAFD-RAG demonstrates consistent performance across all embedding models, with Comprehensiveness sc...

work page 2024
[5]

For each identified entity, extract the following information: - entity_name: Name of the entity, use the same language as input text

Identify all entities. For each identified entity, extract the following information: - entity_name: Name of the entity, use the same language as input text. If English, capitalize the name. - entity_type: One of the following types: [{entity_types}] - entity_description: Comprehensive description of the entity’s attributes and activi- ties Format each en...

work page
[6]

relationship

From the entities identified in step 1, identify all pairs of (source_entity, target_entity) that are clearly related to each other. For each pair of related entities, extract the following information: - source_entity: name of the source entity, as identified in step 1 - target_entity: name of the target entity, as identified in step 1 - relationship_des...

work page
[7]

content_keywords

Identify high-level keywords that summarize the main concepts, themes, or topics of the entire text. These should capture the overarching ideas present in the document. Format the content-level keywords as (“content_keywords”{tuple_delimiter}<high_level_keywords>)

work page
[8]

Use **{record_delimiter}** as the list delimiter

Return output in {language} as a single list of all the entities and relationships identified in steps 1 and 2. Use **{record_delimiter}** as the list delimiter. 30 Published as a conference paper at ICLR 2026

work page 2026
[9]

high_level_keywords

When finished, output {completion_delimiter} A4.3 KEYWORDEXTRACTIONPROMPT This prompt identifies salient concepts and entities from a given query for the purpose of initializing seed nodes during query-aware diffusion. It follows Chen et al. (2025a)’s design without any modification. Prompt 2: Keyword-Extraction in Knowledge Graph Question Answering —Role...

work page 2026
[10]

Review this schema summary and user query thoroughly

work page
[11]

Break the user query into logical user subqueries

work page
[12]

• Target nodes: ALL DESTINATION COLUMNS (with table prefixes) having strong relationships of ANY TYPE ABOVE with the subquery

For each subquery, systematically examine EVERY single node (table.column) in the database schema graph one by one: • Go through each table in the schema summary sequentially • For each table, examine every column within that table • For each table.column node, evaluate its relationship strength with the current subquery • Document your analysis for each ...

work page
[13]

When domain-specific concepts appear in the query, properly map these concepts to the appropriate schema column elements

work page
[14]

Identify the most confident path of schema graph: For each subquery, determine and explicitly state the most confident path that the LLM should follow through the schema graph, using right arrow format (→) with ONLY schema nodes

work page
[15]

table.column

Provide reasoning confidence candidates: For each subquery, provide EXACTLY 2-3 DIVERSE candidates in format [source, target, confidence] where source is a starting node, target is an ending node, and confidence is a float between 0.0-1.0. Each candidate should explore different interpretations, relationship types, or alternative paths. Important: • Schem...

work page 2026
[16]

You can check DDL.csv file with the database’s DDL, along with JSON files that contain the column names, column types, column descriptions, and sample rows for individual tables

If your first query fails, then you should explore the database structure further using the methods below. You can check DDL.csv file with the database’s DDL, along with JSON files that contain the column names, column types, column descriptions, and sample rows for individual tables. You can review the DDL.csv file in each directory, then selectively exa...

work page
[17]

Do not use this action to query INFORMATION_SCHEMA or SHOW DATABASES/TABLES; the schema information is all stored in the /workspace/- database_name folder

You can use SNOWFLAKE_EXEC_SQL to run your SQL queries and interact with the database. Do not use this action to query INFORMATION_SCHEMA or SHOW DATABASES/TABLES; the schema information is all stored in the /workspace/- database_name folder. Refer to this folder whenever you have doubts about the schema. 34 Published as a conference paper at ICLR 2026

work page 2026
[18]

Once it makes sense, consider it resolved

Be prepared to write multiple SQL queries to find the correct answer. Once it makes sense, consider it resolved

work page
[19]

Focus on SQL queries rather than frequently using Bash commands like grep and cat, though they can be used when necessary

work page
[20]

Do not output the same SQL queries repeatedly

If you encounter an SQL error, reconsider the database information and your previous queries, then adjust your SQL accordingly. Do not output the same SQL queries repeatedly

work page
[21]

Once the results are stored in result.csv, make sure the file contains data

Ensure you get valid results, not an empty file. Once the results are stored in result.csv, make sure the file contains data. If it is empty or just contains the table header, it means your SQL query is incorrect

work page
[22]

Save the answer as a CSV and provide the file name, it is usually from the SQL execution result

The final result MUST be a CSV file, not an .sql file, a calculation, an idea, a sentence or merely an intermediate step. Save the answer as a CSV and provide the file name, it is usually from the SQL execution result. TIPS

work page
[23]

For example, for /workspace/DEPS_DEV_V1/DEPS_DEV_V1/ADVISORIES.json, if you want to use it in SQL, you should write DEPS_DEV_V1.DEPS_DEV_V1.ADVISORIES

When referencing table names in Snowflake SQL, you must include both the database_name and schema_name. For example, for /workspace/DEPS_DEV_V1/DEPS_DEV_V1/ADVISORIES.json, if you want to use it in SQL, you should write DEPS_DEV_V1.DEPS_DEV_V1.ADVISORIES

work page
[24]

Do not write SQL queries to retrieve the schema; use the existing schema documents in the folders

work page
[25]

When encountering bugs, carefully analyze and think them through; avoid writing repetitive code

work page
[26]

But don’t use \”, just use “

Column names must be enclosed in quotes. But don’t use \”, just use “. RESPONSE FORMAT For each task input, your response should contain:

work page
[27]

Thought:

One analysis of the task and the current environment, reasoning to determine the next action (prefix “Thought: ”)

work page
[28]

Action:

One action string in the ACTION SPACE (prefix “Action: ”). EXAMPLE INTERACTION Observation: ...(the output of last actions, as provided by the environment and the code output, you don’t need to generate it) Thought: ... Action: ... TASK Please solve this task: {task} A6 TEXT-TO-SQL BASELINES CHASE-SQL for SQLiteWe made several key modifications to adapt C...

work page 2026
[29]

Make sure to change the syntax and functions specifically for snowflake, but do not change the prompt structure and do not omit any examples

work page
[30]

All examples must be executable in snowflake

work page
[31]

Make sure to change all ticks <‘> with double quotes <">

work page
[32]

You can infer the best names for the database and schema if they are not immediately available

All DDL statements specifying the database structure must include statements for creating a database and schema. You can infer the best names for the database and schema if they are not immediately available. Use the same name for the database and the schema. When referencing tables, use the full namespace including the database and schema, like so: datab...

work page 2026

[1] [1]

doi: 10.18653/v1/2024.findings-emnlp.685

Association for Computational Linguistics. doi: 10.18653/v1/2024.findings-emnlp.685. URLhttps://aclanthology.org/2024.findings-emnlp.685/. Parth Sarthi, Salman Abdullah, Aditi Tuli, Shubh Khanna, Anna Goldie, and Christopher D Manning. Raptor: Recursive abstractive processing for tree-organized retrieval. InThe Twelfth International Conference on Learning...

work page doi:10.18653/v1/2024.findings-emnlp.685 2024

[2] [2]

Saba Sturua, Isabelle Mohr, Mohammad Kalim Akram, Michael Günther, Bo Wang, Markus Krimmel, Feng Wang, Georgios Mastrapas, Andreas Koukounas, Nan Wang, et al

doi: 10.1137/080744888. Saba Sturua, Isabelle Mohr, Mohammad Kalim Akram, Michael Günther, Bo Wang, Markus Krimmel, Feng Wang, Georgios Mastrapas, Andreas Koukounas, Nan Wang, et al. jina-embeddings-v3: Multilingual embeddings with task lora.arXiv preprint arXiv:2409.10173, 2024. Heli Sun, Fang He, Jianbin Huang, Yizhou Sun, Yang Li, Chenyu Wang, Liang He...

work page doi:10.1137/080744888 2024

[3] [3]

uses multi-stage retrieval and ego-network aggregation. PathRAG (Chen et al., 2025a) further improves graph-based RAG by retrieving key relational paths rather than redundant neighborhood information, using flow-based pruning to identify reliable paths and strategic prompt organization to enhance LLM responses. However, all aforementioned methods still st...

work page 2013

[4] [4]

All other hyperparameters remain fixed across experiments

(4096-dim), optimized for retrieval tasks; and (iv) GritLM-7B (Muennighoff et al., 2024) (4096-dim), a unified generative-embedding model. All other hyperparameters remain fixed across experiments. Table 7 presents results across five evaluation dimensions. QAFD-RAG demonstrates consistent performance across all embedding models, with Comprehensiveness sc...

work page 2024

[5] [5]

For each identified entity, extract the following information: - entity_name: Name of the entity, use the same language as input text

Identify all entities. For each identified entity, extract the following information: - entity_name: Name of the entity, use the same language as input text. If English, capitalize the name. - entity_type: One of the following types: [{entity_types}] - entity_description: Comprehensive description of the entity’s attributes and activi- ties Format each en...

work page

[6] [6]

relationship

From the entities identified in step 1, identify all pairs of (source_entity, target_entity) that are clearly related to each other. For each pair of related entities, extract the following information: - source_entity: name of the source entity, as identified in step 1 - target_entity: name of the target entity, as identified in step 1 - relationship_des...

work page

[7] [7]

content_keywords

Identify high-level keywords that summarize the main concepts, themes, or topics of the entire text. These should capture the overarching ideas present in the document. Format the content-level keywords as (“content_keywords”{tuple_delimiter}<high_level_keywords>)

work page

[8] [8]

Use **{record_delimiter}** as the list delimiter

Return output in {language} as a single list of all the entities and relationships identified in steps 1 and 2. Use **{record_delimiter}** as the list delimiter. 30 Published as a conference paper at ICLR 2026

work page 2026

[9] [9]

high_level_keywords

When finished, output {completion_delimiter} A4.3 KEYWORDEXTRACTIONPROMPT This prompt identifies salient concepts and entities from a given query for the purpose of initializing seed nodes during query-aware diffusion. It follows Chen et al. (2025a)’s design without any modification. Prompt 2: Keyword-Extraction in Knowledge Graph Question Answering —Role...

work page 2026

[10] [10]

Review this schema summary and user query thoroughly

work page

[11] [11]

Break the user query into logical user subqueries

work page

[12] [12]

• Target nodes: ALL DESTINATION COLUMNS (with table prefixes) having strong relationships of ANY TYPE ABOVE with the subquery

For each subquery, systematically examine EVERY single node (table.column) in the database schema graph one by one: • Go through each table in the schema summary sequentially • For each table, examine every column within that table • For each table.column node, evaluate its relationship strength with the current subquery • Document your analysis for each ...

work page

[13] [13]

When domain-specific concepts appear in the query, properly map these concepts to the appropriate schema column elements

work page

[14] [14]

Identify the most confident path of schema graph: For each subquery, determine and explicitly state the most confident path that the LLM should follow through the schema graph, using right arrow format (→) with ONLY schema nodes

work page

[15] [15]

table.column

Provide reasoning confidence candidates: For each subquery, provide EXACTLY 2-3 DIVERSE candidates in format [source, target, confidence] where source is a starting node, target is an ending node, and confidence is a float between 0.0-1.0. Each candidate should explore different interpretations, relationship types, or alternative paths. Important: • Schem...

work page 2026

[16] [16]

You can check DDL.csv file with the database’s DDL, along with JSON files that contain the column names, column types, column descriptions, and sample rows for individual tables

If your first query fails, then you should explore the database structure further using the methods below. You can check DDL.csv file with the database’s DDL, along with JSON files that contain the column names, column types, column descriptions, and sample rows for individual tables. You can review the DDL.csv file in each directory, then selectively exa...

work page

[17] [17]

Do not use this action to query INFORMATION_SCHEMA or SHOW DATABASES/TABLES; the schema information is all stored in the /workspace/- database_name folder

You can use SNOWFLAKE_EXEC_SQL to run your SQL queries and interact with the database. Do not use this action to query INFORMATION_SCHEMA or SHOW DATABASES/TABLES; the schema information is all stored in the /workspace/- database_name folder. Refer to this folder whenever you have doubts about the schema. 34 Published as a conference paper at ICLR 2026

work page 2026

[18] [18]

Once it makes sense, consider it resolved

Be prepared to write multiple SQL queries to find the correct answer. Once it makes sense, consider it resolved

work page

[19] [19]

Focus on SQL queries rather than frequently using Bash commands like grep and cat, though they can be used when necessary

work page

[20] [20]

Do not output the same SQL queries repeatedly

If you encounter an SQL error, reconsider the database information and your previous queries, then adjust your SQL accordingly. Do not output the same SQL queries repeatedly

work page

[21] [21]

Once the results are stored in result.csv, make sure the file contains data

Ensure you get valid results, not an empty file. Once the results are stored in result.csv, make sure the file contains data. If it is empty or just contains the table header, it means your SQL query is incorrect

work page

[22] [22]

Save the answer as a CSV and provide the file name, it is usually from the SQL execution result

The final result MUST be a CSV file, not an .sql file, a calculation, an idea, a sentence or merely an intermediate step. Save the answer as a CSV and provide the file name, it is usually from the SQL execution result. TIPS

work page

[23] [23]

For example, for /workspace/DEPS_DEV_V1/DEPS_DEV_V1/ADVISORIES.json, if you want to use it in SQL, you should write DEPS_DEV_V1.DEPS_DEV_V1.ADVISORIES

When referencing table names in Snowflake SQL, you must include both the database_name and schema_name. For example, for /workspace/DEPS_DEV_V1/DEPS_DEV_V1/ADVISORIES.json, if you want to use it in SQL, you should write DEPS_DEV_V1.DEPS_DEV_V1.ADVISORIES

work page

[24] [24]

Do not write SQL queries to retrieve the schema; use the existing schema documents in the folders

work page

[25] [25]

When encountering bugs, carefully analyze and think them through; avoid writing repetitive code

work page

[26] [26]

But don’t use \”, just use “

Column names must be enclosed in quotes. But don’t use \”, just use “. RESPONSE FORMAT For each task input, your response should contain:

work page

[27] [27]

Thought:

One analysis of the task and the current environment, reasoning to determine the next action (prefix “Thought: ”)

work page

[28] [28]

Action:

One action string in the ACTION SPACE (prefix “Action: ”). EXAMPLE INTERACTION Observation: ...(the output of last actions, as provided by the environment and the code output, you don’t need to generate it) Thought: ... Action: ... TASK Please solve this task: {task} A6 TEXT-TO-SQL BASELINES CHASE-SQL for SQLiteWe made several key modifications to adapt C...

work page 2026

[29] [29]

Make sure to change the syntax and functions specifically for snowflake, but do not change the prompt structure and do not omit any examples

work page

[30] [30]

All examples must be executable in snowflake

work page

[31] [31]

Make sure to change all ticks <‘> with double quotes <">

work page

[32] [32]

You can infer the best names for the database and schema if they are not immediately available

All DDL statements specifying the database structure must include statements for creating a database and schema. You can infer the best names for the database and schema if they are not immediately available. Use the same name for the database and the schema. When referencing tables, use the full namespace including the database and schema, like so: datab...

work page 2026