Documentation
¶
Overview ¶
Package ragnotify implements agent.DocFinder over an MCP retrieval server (e.g. ragtag) — the concrete half of agentkit's proactive-retrieval seam.
agent owns the neutral machinery (DocFinder + FinderPreparer in agent/finder.go): observe the conversation, gate/cap/dedup, inject pointer notices. This package owns the one thing agent must NOT (invariant #2: agent imports only llm + mcpmgr + stdlib) — talking to an actual retrieval backend. It depends on agent + mcpmgr only.
The whole integration is: point the SAME MCP server at both channels.
- Explicit: advertise the server's search tool to the model normally (the mcpToolDefs bridge). The model calls it when it knows to.
- Passive: wrap that same tool in an MCPFinder and hand it to agent.FinderPreparer. A pointer notice says "doc X looks relevant"; the model fetches the body with the tool it already has.
Reranking is the server's job. This finder passes the conversation text straight to the tool and trusts the ordering/scores it returns — agent's MinScore/MaxHits only trims a pre-ranked list. If ragtag returns raw vector top-k without a reranker, expect noise regardless of MinScore.
Index ¶
Constants ¶
This section is empty.
Variables ¶
This section is empty.
Functions ¶
func MCPFinder ¶
MCPFinder returns an agent.DocFinder backed by an MCP search tool on mgr. serverID + toolName name the ragtag server and its search tool (the same tool you advertise to the model for the explicit channel).
func ParseHits ¶
ParseHits maps a ragtag search result string onto []agent.DocHit. It accepts either a bare JSON array of hits or an object wrapping them under "hits" / "results" / "documents" — the shapes retrieval servers commonly return. A hit missing an id is dropped (the id is the dedup + fetch key). Exported so a host with a different wire shape can test its own adapter against this one.
Types ¶
type Opts ¶
type Opts struct {
// QueryArg is the tool argument that carries the query text. Default "query".
QueryArg string
// ExtraArgs are merged into every call (e.g. {"top_k": 8, "strategy":
// "hybrid"}) — this is where you pin the server's rerank strategy.
ExtraArgs map[string]any
// Join combines the observed entry texts into one query string. Default:
// newline-join (most recent conversation turns as one query).
Join func(texts []string) string
}
Opts configures how the conversation text maps onto an MCP tool call and how the tool's JSON result maps back onto []agent.DocHit. All fields default; a backend with different field names overrides them without touching agent.
Source Files