Skip to content

ARCHITECTURE.md

Latest commit

 

History

History
116 lines (94 loc) · 3.22 KB

ARCHITECTURE.md

File metadata and controls

116 lines (94 loc) · 3.22 KB

ARCHITECTURE

DSGo is a 3-layer framework for building structured LLM applications.

Layers

graph TB
  subgraph "Layer 3: Modules"
    M1[Predict]
    M2[ChainOfThought]
    M3[ReAct]
    M4[Refine]
    M5[BestOfN]
    M6[Program]
    M7[Parallel]
    M8[ProgramOfThought]
    M9[MultiChainComparison]
  end

  subgraph "Layer 2: Core"
    C1[Signature]
    C2[Adapter]
    C3[Prediction]
    C4[History]
    C5[Tools]
    C6[Cache]
  end

  subgraph "Layer 1: Providers"
    P1[OpenAI]
    P2[OpenRouter]
  end

  M1 --> C1
  M2 --> C1
  M3 --> C5

  C2 --> P1
  C2 --> P2

  C1 --> C2
  C2 --> C3
Loading
  • Public API lives in dsgo (see dsgo.go).
  • Implementations live under internal/*.

Links:

Modules

DSGo modules are higher-level behaviors built on top of the core Signature + Adapter + LM primitives.

  • Predict: Basic structured input → output prediction.
  • ChainOfThought: Adds a reasoning step and stores it in Prediction.Rationale (plus any explicit reasoning output fields).
  • ReAct: A native tool-calling loop agent with an auto-injected finish tool and a post-loop extractor fallback to produce signature-valid outputs.
  • Refine: Iteratively improves a prediction when inputs["feedback"] (or a custom refinement field) is provided.
  • BestOfN: Runs a wrapped module N times and selects the best result using a scorer; can parallelize and optionally return all completions.
  • Program: Sequential pipeline; merges previous outputs into the next step's inputs.
  • Parallel: Runs the wrapped module concurrently across expanded inputs; stores per-item outputs in Prediction.Completions.
  • ProgramOfThought: Generates code-first solutions; execution is disabled by default and, when enabled, runs local python3/node under a timeout (no tool loop).
  • MultiChainComparison: Synthesizes a final answer from inputs["completions"] (multiple attempts) and prepends a rationale output.

Request lifecycle

flowchart LR
  A[Inputs] --> B[Adapter formats prompt]
  B --> C[Provider generates]
  C --> D[Adapter parses]
  D --> E[Validation]
  E --> F[Prediction]
Loading

Adapter fallback

graph TD
  A[Raw response] --> B{ChatAdapter}
  B -->|ok| R[Result]
  B -->|fail| C{JSONAdapter}
  C -->|ok| R
  C -->|fail| D{TwoStepAdapter}
  D -->|ok| R
  D -->|fail| X[Error]
Loading

Tool calling (ReAct loop)

flowchart TD
  Q[Inputs] --> M[LM step]
  M -->|tool call(s)| A[Act: tool execution]
  A --> O[Observation(s)]
  O --> M
  M -->|finish(args) or direct answer| F[Final outputs (validated)]
  F -->|if invalid or limits hit| X[Extractor (schema enforced)]
Loading

Notes:

  • DSGo auto-injects a synthetic finish tool that mirrors the output Signature.
  • If the loop can’t produce a valid final output (or exceeds limits), DSGo runs an extraction step over the full trajectory to salvage a schema-valid answer.

Streaming pipeline

flowchart LR
  P[Provider stream] --> B[StreamingBuffer]
  B --> M[Marker filter]
  M --> C[Chunks]
  M --> F[Final parsed Prediction]
Loading