Attach Modes

When an agent installs an IMPP artifact, it needs a strategy for incorporating that knowledge at runtime. IMPP defines four attach modes that control how artifact content is delivered to the receiving model.

Overview

prepend_policy

The simplest mode. The full artifact content is prepended to the agent's system prompt on every invocation.

import impp

artifact = impp.load("defi-risk-assessment@v2.1")
artifact.attach(mode="prepend_policy")

Tradeoffs:

• Lowest latency (no retrieval step)
• Uses context window on every call, even when irrelevant
• Best for small, always-relevant artifacts (risk policies, compliance rules)

few_shot_examples

Artifact sections are formatted as input/output pairs and injected into the prompt as few-shot demonstrations.

artifact.attach(mode="few_shot_examples", max_examples=5)

Tradeoffs:

• Strong calibration effect — the model mimics demonstrated patterns
• Requires artifacts with clearly separable sections
• Token cost scales with example count

retrieval_only

The artifact content is indexed locally. At runtime, the agent's query is matched against artifact sections, and only relevant sections are injected into context.

artifact.attach(mode="retrieval_only", top_k=3)

Tradeoffs:

• Scales to large artifacts without exhausting context
• Adds retrieval latency (~50-200ms depending on index size)
• Relevance depends on query quality

tool_policy

Artifact knowledge is exposed as tool descriptions or function schemas. The model decides when to invoke the knowledge based on task context.

artifact.attach(mode="tool_policy")

Tradeoffs:

• Most flexible — model decides when knowledge is relevant
• Requires a model that supports tool/function calling
• Higher autonomy, lower predictability

Choosing a Mode

The right mode depends on artifact size, how often the knowledge is relevant, and the agent's architecture:

• Always relevant + small → prepend_policy
• Calibration / demonstration → few_shot_examples
• Large reference corpus → retrieval_only
• Agentic decision-making → tool_policy

Modes can be changed at runtime without re-publishing. The artifact content is the same — only the delivery strategy differs.