Use cases & endpoints
Financial data for LlamaIndex — retrieval over filings
LlamaIndex work on financial documents usually stalls in the same place: ingestion. Filings arrive as HTML with inconsistent structure, transcripts as undifferentiated text, central bank releases as PDFs, and the chunking strategy that works for one corpus destroys the meaning of another. Splitting a financial statement mid-table is worse than not indexing it.
Hindcast provides these documents already chunked with their structure intact — filing sections, transcript speaker turns, statement paragraphs — and with the metadata that retrieval needs to be useful: entity, document type, publication timestamp, and a link to the original.
That metadata is what makes filtered retrieval possible. Restricting a query to documents published before a date, or to one company's filings, or to Q&A rather than prepared remarks, turns a vague semantic search into a precise one, and gives the generation step something citable to work from.
hindcast — session
previewhindcast> retrieve filing passages published before a given date
→ tool: filings_search { query:"...", entity:"...", published_before:"..." }
✓ chunked passages with section, document and publication date
source: preview — endpoints open to the waitlist first
What's covered
- Documents pre-chunked along their real structure, not fixed lengths
- Metadata for filtered retrieval: entity, type, date, section
- Publication timestamps enabling as-of retrieval
- Filings, transcripts and central bank text in one schema
- Links back to source documents for citation in generated answers
- Entity identifiers shared with the numeric datasets
Join the waitlist for early access.
Questions
Can I bring my own embeddings and vector store?
Yes. The value here is the corpus and its metadata; where you embed and store it is your decision, and nothing in the design assumes a particular vector database.
Why does publication-date filtering matter for RAG?
Because a retrieval pipeline that pulls a document published after the date being analyzed will produce an answer that looks well-sourced and is anachronistic. Date filters are the retrieval equivalent of point-in-time discipline.
Related
Join the waitlist. Vote the roadmap. First in, first served.