solvi guide¶
The reference behind the documentation's three parts: Using solvi (the ready systems and how you set them up),
Building blocks (the low level, solvi.core, they are made of) and Experimental (solvi.experimental). On the
site each chapter is a page of its part; here they follow in that order. Where a question needs judgement, a model
proposes and solvi's checks decide; the model is whichever you have — an LLM through solvi.models.llm, a decision
service through solvi.models.systemone, or a local checkpoint such as solvi-base (solvi.models.decider) for offline
or cheap cases (see the model proposes). For a two-minute overview, see
the README; for agents and knowledge, Using solvi: agents and knowledge; for advice drawn
from building on solvi, see best practices. Every measured number in this guide names its source:
a script in benchmarks/ or an example, which you can re-run from this repository, or the model card of a published
model (solvi-base, solvi-large, solvi-large-long, extract-base, extract-receipts).
Contents:
Using solvi — the ready systems and what you write to use them: solvi.build, the catalog and its questions, models
from solvi.models, System and Response, solvi.Guard, serving, the command line. Agents and knowledge
(solvi.Agent, solvi.Knowledge) have a page of their own.
- Quick start: solvi.build
- Concepts
- Installation
- The catalog
- Questions and answer types
- Types, questions and model decisions
- Asking: System and Response
- Questions without a rule: fit, learn_rule, teach
- Guarding an agent's tool calls
- Serving: HTTP, MCP and System One
- Text in: from a message to a question
- Checking a catalog: solvi check
- Printing results: solvi.show
- Command line
- Guarantees and limitations
Building blocks — the low level, solvi.core, that the ready systems are made of — to rebuild a part or write your
own. Building blocks has one reference per extension point.
- How the strategist plans a flow
- Confidence, calibration and abstention
- The trace and verification
- Grounded decisions: provenance, audit and safeguards
- A model that writes: generation, agreement and the re-ask loop
- Who answers: System 1, the slow path or a person (solvi.core.dispatch)
- Knowledge: what a system learned, and from whom
- System 1 and System 2 on a game: the Pokémon world map
- Extracting fields from documents
- Versions of a document and their impact: solvi.core.versions
- Specialists: a model proposes, code checks, the store keeps the run
- Explaining a decision: the smallest change that flips it
Experimental — solvi.experimental: works and is tested, no measured gain yet; its API may change, and each piece
graduates or is removed by 1.2 (Experimental lists what each one is missing).