Financial statement extraction
Retrieve tables, amounts and periods from reports. Validate currency, scale and the relationship to the source page.
Industries / 10 / Services and public sector
Banking and financial services
Financial statement extraction, news sentiment analysis and internal policy search. Connect models to your organisation’s tables, documents and text.
Build structured outputs and sourced answers for analysis and operations teams.
Documents, knowledge, and automation
Select a task with defined data and review criteria.
Retrieve tables, amounts and periods from reports. Validate currency, scale and the relationship to the source page.
Classify English financial text as positive, negative or neutral. Use labels as one signal within your analysis.
Query procedures and policies using authorised sources. Filter by department, validity and team permissions.
Organise submitted documents and extract workflow fields. Your application checks missing information and prepares the next task.
With QDivZero / Banking and financial services
Convert pages and tables into reviewable data or classify text with a financial model. Connect results to your systems; Retrieval supplies context for questions about policies and procedures.
Open-weight / Banking and financial services
FinBERT for English financial sentiment; Granite Docling for document structure; embeddings and Qwen for retrieving and querying policies. Each covers a different task.
Financial sentiment · English
Classify English financial text as positive, negative or neutral.
Hugging Face referenceView modelDocument conversion
For converting financial report pages and tables before extracting fields and validating figures.
Hugging Face referenceView modelEmbeddings
For retrieving policies and procedures by department, permissions and document version.
View modelText and vision
For answering with internal policy context and preparing reviewable outputs from your documents.
View modelBanking and financial services
ProsusAI/finbert classifies English financial text into positive, negative and neutral sentiment. You can evaluate it with news or report passages. It provides a text signal for analysis; it does not calculate creditworthiness, credit risk or an investment decision.
Process pages with a document model and define the required data schema. Preserve headers, period, currency and scale alongside page references. Validate row and column relationships before consolidating amounts across documents or financial years.
Index sources with department, version and permissions, then retrieve passages relevant to each query. An LLM can prepare a referenced answer. Your application controls access and connects tools when it needs current operational data.
Define workflow-related categories and fields, such as document type, date or identifier. Evaluate incomplete formats and unknown types. The application validates results and prepares the next step; retain originals so the team can review classification or extracted data.
Separate tasks: a document model retrieves structure and tables; FinBERT supplies English sentiment; embeddings retrieve passages; an LLM prepares contextual answers. You can combine components without assuming one model performs numerical analysis, search and classification equally well.
Separate document preparation, retrieval, inference and tool integration. Estimate volume, concurrency and context for each workflow: classifying requests differs from analysing long contracts. Compare configurations with real examples and include result validation and review.
Your integration can expose authorised queries through tools using those systems’ APIs. The model requests a tool and your application validates and executes the call. Limit accessible data and distinguish queries from actions that change records.
You can deploy compatible private or fine-tuned models and connect them to your application. Evaluate formats, context and instruction following with workflow examples. If the goal is querying changing documentation, also compare a Retrieval solution before deciding how to adapt the model.
Define workflow-specific validation: required fields, consistency of amounts, references and permitted actions. Test incomplete inputs and ambiguous cases. Your application uses these criteria to proceed or request review; a well-written answer alone does not validate a value or operation.
Compare extraction, references and formats using your organisation’s documents before expanding the workflow.