DataLens Skill encompasses a query handler for natural language questions.

Query history contains Query records representing individual answered questions stored in the database. The Frontend uses the Query History entity for the analysis view.

Requirement category with unmet compliance; involves iterative process for improving query accuracy. The plan includes query refinement to reuse the Verifier Agent for regenerating SQL queries upon failure.

Query Router currently reads extracted data from DuckDB. After migration, Query Router will read extracted data from PostgreSQL schemas.

Intelligent consolidation improves the query success rate from 70% to over 95%.

DataLens requires query tracking middleware to implement audit logging of queries and users for regulatory compliance. DataLens requires adding query tracking middleware for audit logs and compliance. DataLens requires query tracking middleware to track user queries

DataLens Agent Mode implements the query-data skill to route natural language and SQL queries for data analysis.

QueryComplexityClassifier is a component of the Multi-Stage Text-to-SQL Architecture used to classify query complexity and route models. Multi-Stage Text-to-SQL Architecture realizes the QueryComplexityClassifier use case for determining question complexity and model routing.

QueryDataSkill produces SkillResult when executing natural language queries via Text-to-SQL. DuckDBService provides database connections and query execution needed by QueryDataSkill. QuestionRouter routes classification results to QueryDataSkill for SQL query execution. TextToSQLService enables QueryDataSkill to convert natural language queries to SQL queries.

The backend's QuestionRouter in backend/app/services/question_router.py classifies questions and routes queries to DuckDB SQL backend or Qdrant semantic search. It lazily initializes QdrantService to avoid startup delays, managing structured, textual, and hybrid query paths. It works with TextToSQLService and DuckDBService for SQL generation and execution, and handles query classification and routing within the API layer. QdrantService initialization was changed to lazy loading in the QuestionRouter to prevent startup timeouts QuestionRouter depends on Qdrant, but now falls back gracefully if Qdrant is unavailable Question router routes textual queries to Qdrant semantic search service. The /ask endpoint uses the Question router to classify and route user questions. The Question router capability is validated by the test_question_router test case.

question_router.py is updated to use pg_data_service.py for reading extracted data in place of DuckDBService. question_router.py uses translate_duckdb_to_pg to convert DuckDB-style SQL to PostgreSQL-compatible SQL before execution. The question_router.py module depends on the FindingsGenerator to generate findings from query results. The question_router.py module was modified to use PgDataService instead of DuckDBService for query operations. The LLM Prompt Injection use case modifies prompts in question_router.py to include language parameter for answer synthesis. question_router.py uses FindingsGenerator to generate findings from query results

The QuestionRouter class is defined within backend/app/services/question_router.py. The QuestionRouter class uses the TEXT-TO-SQL Service for SQL query generation. The QuestionRouter class uses the DUCKDB Service for executing SQL queries and retrieving data. The QuestionRouter class uses the QDRANT SERVICE for semantic search over vector data. The QuestionRouter class is used by the API LAYER including the backend/app/api/analysis.py endpoint for processing queries. QuestionRouter.route() is the main method within QuestionRouter that orchestrates query execution. The route method is part of the QuestionRouter class.

QuestionRouter integration works with FindingsGenerator logic in the agent architecture to process and generate findings. IronClaw Agent depends on QuestionRouter.route() to route queries and execute them correctly as part of the async processing pipeline.

DataLens team started with 100K/200K tokens, consumed 147K tokens and continuously evaluated usage versus quality between Qwen and Sonnet models, ultimately keeping the Sonnet model for quality.

vLLM model deployed on elin with 14B parameters, optimized for GPU inference, using a 4-bit quantization to fit in 10GB VRAM. Qwen2.5-Coder-14B-AWQ is a part of the DataLens DS-STAR Implementation Plan as the deployed vLLM model. The vLLM component uses the Qwen2.5-Coder-14B-AWQ model version.

The Multi-Stage Text-to-SQL Architecture uses Qwen3 for complex query handling, table selection, and answer synthesis. TableReranker uses the Qwen3 model to re-rank candidate tables and select the most relevant ones for query answering. AnswerSynthesizer leverages Qwen3 to synthesize human-readable answers in Danish or English from SQL query results. The Multi-Stage Architecture uses the Qwen3 model for schema selection and answer synthesis phases. The Data Discovery feature architecture uses Qwen3 with large context for table selection and Arctic with smaller context for SQL generation. The Data Discovery feature architecture uses Qwen3 with large context for table selection and Arctic with smaller context for SQL generation. The Data Discovery System requires the Qwen3 table selection capability to identify relevant tables before SQL generation. SQL extraction regex fix addresses the problem in Qwen3 response format where the SQL query was not properly captured due to missing newline before closing backticks. Multi-Stage Text-to-SQL Architecture uses Qwen3 for schema comprehension and complex SQL tasks. The Discovery Service uses Qwen3 LLM for table selection in the intelligent table discovery process. The Data Discovery Architecture uses Qwen3 for table selection with a large context window. The Qwen3 response format required the SQL extraction regex to be fixed to handle SQL code blocks without a newline before closing backticks. The Live Backend processes Qwen3 multi-block XML responses for SQL extraction.

RAG Agent generates answers using the qwen3-coder-next LLM via Ollama The Ollama API provides access to the qwen3-coder-next model. qwen3-coder-next model is deployed and running on the elin server before the deployment of SQLCoder-7B. SQLCoder-7B replaced qwen3-coder-next as the default model for Text-to-SQL queries to improve speed from 40-50s to 2-3s per inference. SQLCoder-7B is proposed to be used for Text-to-SQL queries, while qwen3-coder-next may still be used for summary generation tasks. SQLCoder-7B is deployed as a replacement for qwen3-coder-next to achieve significant speed improvements in DataLens queries. qwen3-coder-next continues to be considered as a fallback model for AI summary generation when SQLCoder-7B may fail in that role.

Refers to Retrieve-Augment-Generate technique used in backend for document retrieval and question answering. Document RAG provides a RAG query interface to perform semantic searches over documents.

The plan includes RAGAgent to query unstructured text and synthesize answers with vLLM. RAG Agent uses Qdrant for semantic vector search of documents RAG Agent uses Ollama embeddings via nomic-embed-text for document retrieval RAG Agent generates answers using the qwen3-coder-next LLM via Ollama RAGAgent corresponds to the rag agent implementation RAGAgent corresponds to the rag agent implementation RAG Agent uses Qdrant for semantic vector search of documents RAG Agent uses Ollama embeddings via nomic-embed-text for document retrieval RAG Agent generates answers using the qwen3-coder-next LLM via Ollama RAGAgent corresponds to the rag agent implementation RAGAgent corresponds to the rag agent implementation RAG Agent uses Qdrant for semantic vector search of documents RAG Agent uses Ollama embeddings via nomic-embed-text for document retrieval RAG Agent generates answers using the qwen3-coder-next LLM via Ollama RAGAgent corresponds to the rag agent implementation RAGAgent corresponds to the rag agent implementation The RAGAgent is part of Document RAG to perform retrieval and answer synthesis. RAG Agent integrates with Qdrant for vector storage and semantic search. RAG Agent uses Ollama to generate embeddings and answers based on semantic search results. RAG Agent employs the nomic-embed-text embedding model for creating document embeddings. RAG Agent realizes the Natural Language Querying capability by semantic search and answer generation over documents. Lightweight Implementation uses RAG Agent for document question answering without heavy dependencies. RAGAgent optionally searches unstructured text in the data analysis process. RAGAgent uses LlamaIndex to chunk PDFs and generate embeddings. RAGAgent stores embeddings in the Qdrant vector database for retrieval.

The plan optionally uses RAPIDS cuDF for accelerating large dataframe computations on GPU. The plan optionally uses RAPIDS cuDF for accelerating large dataframe processing on elin GPU. RAPIDS cuDF is optionally used for large dataframe acceleration alongside DuckDB.

Acceptance document with unconditional delivery, contents unspecified in messages; no update provided. The DataLens platform includes the README.md file documenting the project overview.

The Real-World Validation System uses datasets sourced from Municipal Finance Data Research for testing. DataLens Development includes an automated validation framework that runs real-world dataset tests.

Endpoint for generating project recommendations, not yet implemented.

A formal requirement model for API requests related to recommendations, with compliance issues noted.