Docling was selected over TableGPT2 for the SVGV Budget Analysis Project due to better suitability and accuracy for Excel extraction without GPU requirements. TableGPT2 was evaluated but rejected in favor of Docling for Excel file extraction due to older technology and less fitting use case.

TableIndexService uses QdrantService to build semantic search indices for database tables. QdrantService supports TableIndexService by providing vector collections for semantic table search indexes. TableMatch represents individual tables that are part of TableIndexService semantic search indexes.

Service class in backend/app/services/table_index.py that manages matching database tables to queries. TableMatch represents individual tables that are part of TableIndexService semantic search indexes.

TableReranker uses the Qwen3 model to re-rank candidate tables and select the most relevant ones for query answering. Multi-Stage Text-to-SQL Architecture realizes the TableReranker use case for LLM-based candidate table re-ranking.

Extractor Agents extract tables from PDFs.

The DOCX extractor embeds extracted tables as JSON within text chunks instead of storing them as separate DuckDB tables. The Docling extraction system uses a business rule that tables extracted from documents are embedded as JSON within semantic chunks.

The Frontend uses Tailwind for styling. The DiscoveryFlow component uses Tailwind CSS for styling and responsive design.

The Frontend utilizes Tailwind CSS The npm dev dependency tailwind-merge depends on tailwindcss as part of the frontend tooling.

The npm dev dependency tailwind-merge depends on tailwindcss as part of the frontend tooling.

IronClaw is constrained by the use of TEE credential vault for security.

Telegram bot operates on the Telegram platform. DataLens Skill integrates with Telegram to provide user interface access. DataLens Skill integrates with Telegram to receive file uploads and user queries. DataLens Skill handles user requests via Telegram interface on theo.

DataLens Skill implements a Telegram bot for interaction. Telegram bot operates on the Telegram platform. DataLens Skill capability includes integration of a Telegram bot for file upload and user interaction. The DataLens Orchestrator uses Telegram UI as the front-end communication channel for users. DataLens Skill involves building a Telegram bot for user interaction.

DataLens Skill includes development of a Telegram upload handler to interface with elin services.

The Data Consolidation capability creates TEMP VIEWs to unify related tables into a single schema for Arctic SQL generation.

DataLens is focused on data analysis projects, while TerrainLens is handled by the main agent, indicating separation of concerns between the two agents. The DataLens Master Implementation Plan depends on TerrainLens usage for GPU resources allocation during Ollama calls. DataLens agent focuses on data analysis and must stay focused on its project, as TerrainLens is handled by the main agent and is a separate concern. DataLens focuses on data analysis and explicitly excludes handling TerrainLens, which is managed by the main agent.

Valid credentials for admin user used in testing discovery, agent access, and deployment verification. Account has Danish language preference, enabling full feature usage.

DataLens consults the Test Engineer when needing validation or feedback on testing concerns. Test engineering practices are employed in the DataLens platform backend development. The DataLens platform backend includes the test-engineering skill. The DataLens platform backend integrates the test-engineering skill.

Issues with test selectors, URL configurations, and database migrations caused test failures. Corrections in selectors and migration application fixed the problems, enabling successful E2E tests.

The GPU-first document extraction implementation is validated by the test suite 'test_docling_extractors.py'. The E2E Test Suite includes Playwright tests for the Data Discovery feature. The E2E Test Suite uses real SVGV budget data for end-to-end validation of the Data Discovery feature.

Script for running comprehensive automated tests on the discovery workflow.

The entire batch upload pipeline is end-to-end validated by the test_batch_pipeline_e2e test case.

test_data_анкеты table is stored in DuckDB (analytics.db)

test_data_шаблон table is stored in DuckDB (analytics.db)

The DOCX extractor capability is validated by the test_extractors test case. The PPTX extractor capability is validated by the test_extractors test case.