hot.dev/hot-ai / ::ai::rag

Functions

AgentMemory alias

Alias of ::ai::memory/AgentMemory

ChatOptions alias

Alias of ::ai::chat/ChatOptions

MemoryCapsule alias

Alias of ::ai::memory/MemoryCapsule

Role alias

Alias of ::ai::message/Role

Thread alias

Alias of ::ai::memory/Thread

ask

fn (mem: ::ai::memory/AgentMemory, question: Str, chat-opts: ::ai::chat/ChatOptions, opts: Map?): Map

Search memory for relevant context, then generate an AI answer. Returns {answer, sources}.

ask-with-thread

fn (mem: ::ai::memory/AgentMemory, t: ::ai::memory/Thread, question: Str, chat-opts: ::ai::chat/ChatOptions, opts: Map?): Map

Like ask, but includes thread history for multi-turn context. Appends question and answer to the thread automatically.

build-context

fn (mem: ::ai::memory/AgentMemory, query: Str, opts: Map?): Str

Search session memory and KB, format as a context string for AI prompts. Options: limit (default 10), min-score (default 0.5).

build-context-safe

fn (mem: ::ai::memory/AgentMemory, query: Str, opts: Map?): Str

Best-effort build-context. Any failure (no embedding provider configured, transient store error, etc.) degrades to "" so callers can still answer ungrounded rather than aborting the whole reply.

Use this in interactive request paths and demos / CI / offline tutorials where an embedding backend may not be wired. Prefer build-context directly in code paths that should propagate errors.

build-hybrid-context

fn (mem: ::ai::memory/AgentMemory, query: Str, opts: Map?): Str

Build a citation-preserving context string from hybrid vector + graph retrieval.

build-rag-prompt

fn (history-text: Str, context: Str, question: Str): Str

Compose a RAG prompt from optional history, optional context, and a question.

compact-window

fn (mem: ::ai::memory/AgentMemory, chat-opts: ::ai::chat/ChatOptions, opts: Map?): ::ai::memory/MemoryCapsule

Summarize a recent memory window into a durable MemoryCapsule and store it in the KB. Options: hours, limit, id, instruction, metadata, emit-stream.

compact-window-safe

fn (mem: ::ai::memory/AgentMemory, chat-opts: ::ai::chat/ChatOptions, opts: Map?): Map

Like compact-window, returning {ok, capsule, error} and emitting compaction errors when requested.

format-kb-results

fn (results: Vec): Vec

Format a Vec of KB search results as [KB] content lines.

format-session-results

fn (results: Vec): Vec

Format a Vec of store search results (with v.sender-name, v.content) as [name]: content lines.

format-thread-lines

fn (history: Vec): Str

Format a Vec of ChatMessages as role: content lines.

ns alias

Alias of ::ai::rag/

Retrieve-Augment-Generate: compose memory search with AI generation. Accepts chat-fn parameter for provider-free AI calls.

passes-role-filter

fn (entry: Map, include-roles: Vec?, exclude-roles: Vec?): Bool

Apply include-roles / exclude-roles semantics:

• When include-roles is non-empty, the entry must have a role in that list. Untagged entries are dropped.
• exclude-roles is a blacklist applied next. Tagged matches are dropped; untagged entries pass.
• With neither set, every entry passes.

recent-text

fn (mem: ::ai::memory/AgentMemory, opts: Map?): Str

Format recent session memory as bulleted - <content> lines, oldest first.

Convenience over ::ai::memory/recent for synthesis prompts that want the raw window as text rather than a finished summary. Errors degrade to "" so missing stores / providers don't crash callers.

Options:

• limit — number of records to include (default 100, matches ::ai::memory/recent).
• hours — recency window in hours (default 24).
• include-roles — Vec<Str>?. When set, only records whose metadata.role is in this list are kept. Untagged records are dropped.
• exclude-roles — Vec<Str>?. When set, records whose metadata.role is in this list are dropped. Untagged records pass.

The role filters compose with limit: filtering happens after recent, so limit bounds the window, not the post-filter count. For synthesis prompts that want only user turns + task records, pass exclude-roles: ["assistant"].

role-in

fn (roles: Vec, role: Str): Bool

True when role equals any element of roles.

role-of

fn (entry: Map): Str

Read the role tag from a ::ai::memory/recent entry's metadata. Returns "" when unset so role filters never accidentally drop untagged records (callers asking to exclude assistant shouldn't also exclude legacy un-roled writes).

summarize

fn (mem: ::ai::memory/AgentMemory, chat-opts: ::ai::chat/ChatOptions, opts: Map?): Str

Summarize recent messages from session memory using AI. Options: hours, limit, instruction.

with-memory-block

fn (base: Str, context: Str, heading: Str?): Str

Compose a system prompt by appending a memory-context section to base.

Returns base unchanged when context is empty (or whitespace only), so callers can pass the result of build-context / build-context-safe directly without branching. When context is non-empty, the result is ${base}\n\n${heading}\n${context}. heading defaults to ## Memory.