Inference Primitives¶

In addition to generative LLMs, Daimon manages two high-throughput deterministic and calibrated inference primitives: NER and Decision Models.

These components allow Daimon to support hybrid domain pipelines (such as medical coding and automated entity disambiguation) without burning expensive generative LLM tokens on span extraction or leaf-node classification.

Component Categories¶

Category	Type Prefix	Description	Primary Drivers
NER	`ner/*`	High-throughput entity extraction, span detection, character offsets, and confidence scoring	`ner/http`, `ner/onnx`
Decision	`decision/*`	Non-autoregressive calibrated discriminative models for candidate ranking and statement verification	`decision/http`, `decision/jev`

End-to-End Pipeline Architecture¶

By combining NER, Vector Stores, Knowledge Graphs, and Decision models, Daimon executes deterministic extraction and disambiguation pipelines:

Raw Unstructured Text
       │
       ▼
[POST /v1/ner/clinical-ner/extract]
       │
       ├─► Returns Entities: ["right knee pain", "mild asthma"]
       │
       ▼
[POST /v1/memory/icd10/query] (Vector Search)
       │
       ├─► Returns Category Nodes: [M25.5, J45.20]
       │
       ▼
[POST /v1/graph/icd10-graph/cypher] (Knowledge Graph)
       │
       ├─► Expands Sub-codes for M25.5: [M25.561, M25.562, M25.569]
       │
       ▼
[POST /v1/decision/verifier/choose]
       │
       ├─ State: "Patient presents with right knee pain..."
       ├─ Question: "Which code accurately describes laterality?"
       ├─ Choices: ["M25.561", "M25.562", "M25.569"]
       │
       ▼
Final Output: "M25.561" (Confidence: 0.97)

Agentic Loop & Tool Synthesis¶

When an NER or Decision component is declared in config.yaml, Daimon registers corresponding tools in the sidecar's tool catalogue. Any configured LLM can call these tools directly during conversational loops:

{name}_extract: Extract spans and entities from text.
{name}_choose: Select the highest-probability choice from candidate options.
{name}_verify: Evaluate calibrated truth probability for a boolean assertion.