Patra
Processing receipt
What was done to a document, by which operation and version, how faithful the result is, digests of what went in and came out, where it ran and whether the document was kept.
Produced by the Patra command-line tool
Home / Technical capabilities
Technical capabilities
Built from the evidence layer upward.
Our systems return records you can check — what was done, what was searched, what supports an answer and what actually ran — on an engineering stack chosen part by part.
The records our systems return
Each one exists in our own systems today. None is offered as a public API yet; they are described here so you can judge the design.
BanyanGraph
Point-in-time retrieval receipt
What was searched, as of which date, limited to what was in force then and ordered by authority — with a trace identifier and a watermark of the corpus it came from.
In our research service; not yet reachable by customers In development
Curator
Evidence packet
The sources behind an answer, with each cited statement marked as supported by them, contradicted by them or going beyond them — sealed so a later change can be detected.
In our research stack; Curator's interface is in development In development
BanyanGraph and Curator
Execution declaration
Which components actually ran, which were missing and why — so neither an answer nor an invoice claims more than ran.
Returned by our research API since July 2026
A complete intelligence chain
Ingestion, structural parsing, enrichment, embeddings, ranking, synthesis, citations and graph relationships are treated as one connected system. Historical news, events and time-series financial context extend the knowledge layer beyond static documents.
- Ingestion
- Structural parsing
- Enrichment
- Embeddings
- Ranking
- Synthesis with citations
- Graph relationships
—
What sits underneath
Vector, graph, relational, time-series and object stores each have one job; optional infrastructure starts only when it is needed, costs are governed, and failures stay visible rather than becoming silent fallbacks. None of it is presented as a universal dependency or as proof of hosted availability.
—
Models, chosen as parts of a system
Open-weight families — Llama, Qwen, Gemma and Nemotron among them — with specialist embedding and reranking models, selected through ongoing experimentation and benchmarking, including during our NVIDIA Inception GPU access. Licence, cost, latency, provenance and failure behaviour are weighed together.
—
A practical goal
We are building cost-conscious systems so smaller firms and larger BFSI organisations can adopt the capabilities their work needs. Current product stages are listed on the capabilities page.
Discuss the technical contract.
If you are weighing how one of these records would fit your systems, tell us what you need to verify. We will walk through the design, and be plain about what exists.