Package com.tnsai.rag

Class RetrievedContextRenderer

java.lang.Object

com.tnsai.rag.RetrievedContextRenderer

public final class RetrievedContextRenderer
extends Object

Renders retrieved documents behind a single, explicit untrusted-context boundary.

Document content and provenance are data controlled by an external knowledge source. This renderer XML-escapes both before placing them inside the <tnsai-memory> fence, preventing retrieved text from terminating the boundary or introducing sibling prompt sections. The original user message is kept outside the fence and remains byte-for-byte unchanged.

Field Summary

Fields

Modifier and Type	Field	Description
static final String	UNTRUSTED_CONTEXT_DISCLOSURE	System-level instruction shown exactly once whenever retrieved evidence is rendered.

Method Summary

Modifier and Type	Method	Description
static String	fence(String context)	Fences an already assembled action context as untrusted data.
static String	render(String userMessage, ChatKnowledgeContext.Turn turn, RetrievalResult result, int contextWindow)	Renders a chat turn's knowledge behind one boundary: the blocks declared before the first SEARCH entry, the SEARCH evidence within the context window, then the blocks declared after it.
static String	render(String userMessage, RetrievalResult result, int contextWindow)	Renders canonical retrieval evidence within the configured context window.
static String	render(String userMessage, List<SearchResult> results)	Renders search results for a single outbound LLM request.
static List<KnowledgeUsage>	searchUsage(List<String> sources, RetrievalResult result, int contextWindow)	Says what each SEARCH source gave when result is rendered within contextWindow: the evidence that fits, as the render methods cut it, attributed to its source.
static String	withUntrustedContextPolicy(String systemPrompt)	Adds the canonical untrusted-context policy to a request system prompt.

Methods inherited from class java.lang.Object

equals, getClass, hashCode, notify, notifyAll, toString, wait, wait, wait

Field Details

UNTRUSTED_CONTEXT_DISCLOSURE

public static final String UNTRUSTED_CONTEXT_DISCLOSURE

System-level instruction shown exactly once whenever retrieved evidence is rendered.

See Also:

• Constant Field Values

Method Details

render

public static String render(String userMessage,
 List<SearchResult> results)

Renders search results for a single outbound LLM request.

Parameters:

userMessage - the original user message

results - retrieved evidence; may be null or empty

Returns:

a fenced request, or the original message when no evidence can be rendered

render

public static String render(String userMessage,
 RetrievalResult result,
 int contextWindow)

Renders canonical retrieval evidence within the configured context window.

Parameters:

userMessage - original user message

result - canonical retrieval result

contextWindow - maximum estimated tokens devoted to evidence

Returns:

fenced evidence followed by the unchanged user message

render

public static String render(String userMessage,
 ChatKnowledgeContext.Turn turn,
 RetrievalResult result,
 int contextWindow)

Renders a chat turn's knowledge behind one boundary: the blocks declared before the first SEARCH entry, the SEARCH evidence within the context window, then the blocks declared after it.

Each block is escaped as a whole, as fence(String) does, because it was assembled before rendering.

Parameters:

userMessage - original user message

turn - the turn's DIRECT blocks

result - SEARCH evidence, or null when nothing was searched

contextWindow - maximum estimated tokens devoted to the evidence

Returns:

fenced knowledge followed by the unchanged user message, or the original message when nothing can be rendered

Since:

0.18.0

searchUsage

public static List<KnowledgeUsage> searchUsage(List<String> sources,
 RetrievalResult result,
 int contextWindow)

Says what each SEARCH source gave when result is rendered within contextWindow: the evidence that fits, as the render methods cut it, attributed to its source. With one source, every fitted item is that source's; with several, each item goes to the source its provenance names. Tokens are the sum of each fitted item's estimate.

Parameters:

sources - the SEARCH sources in declaration order, one row each; when empty, one row per provenance source, in the order the evidence first names it

result - the search result, or null when nothing was searched

contextWindow - maximum estimated tokens devoted to the evidence

Returns:

one SEARCH row per source

Since:

0.18.0

withUntrustedContextPolicy

public static String withUntrustedContextPolicy(String systemPrompt)

Adds the canonical untrusted-context policy to a request system prompt.

Parameters:

systemPrompt - the request's base system prompt

Returns:

the prompt containing the policy exactly once

fence

public static String fence(String context)

Fences an already assembled action context as untrusted data.

Action retrieval supports caller-defined context templates, so the original document boundaries are no longer available at LLM dispatch. Escaping the complete assembled value preserves that formatting as data while preventing it from closing the canonical memory boundary.

Parameters:

context - assembled retrieved context

Returns:

escaped context inside the canonical untrusted boundary