1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
//! Handler for the `mnem_retrieve` MCP tool.
//!
//! Extracted from `tools.rs` in R3; body unchanged.
use std::path::Path;
use super::super::{MAX_RERANK_TOP_K, MAX_RETRIEVE_LIMIT, MAX_VECTOR_CAP};
use crate::server::Server;
use anyhow::{Result, anyhow, bail};
use mnem_core::codec::json_to_ipld;
use mnem_core::index::PropPredicate;
use mnem_core::repo::ReadonlyRepo;
use serde_json::Value;
// ============================================================
// mnem_retrieve
// ============================================================
pub(in crate::tools) fn retrieve(server: &mut Server, args: Value) -> Result<String> {
let repo_path = server.repo_path().to_path_buf();
let allow_labels = server.allow_labels;
let repo = server.load_repo()?;
retrieve_impl(repo, &repo_path, allow_labels, args)
}
pub(super) fn retrieve_impl(
repo: ReadonlyRepo,
repo_path: &Path,
allow_labels: bool,
args: Value,
) -> Result<String> {
#[cfg(not(feature = "summarize"))]
let _ = repo_path;
// `label` gated behind `MNEM_BENCH`. Off by default: the filter is
// silently ignored so retrieve runs unscoped (parity with
// GET/POST /v1/retrieve in mnem http).
let mut r = repo.retrieve();
if allow_labels && let Some(label) = args.get("label").and_then(Value::as_str) {
r = r.label(label);
}
if let Some(Value::Object(map)) = args.get("where")
&& let Some((k, v)) = map.iter().next()
{
let ipld = json_to_ipld(v)?;
r = r.where_prop(k, PropPredicate::Eq(ipld));
}
let text_arg = args.get("text").and_then(Value::as_str).map(str::to_string);
let vector_arg = args.get("vector").and_then(Value::as_object).cloned();
if let Some(ref text) = text_arg {
r = r.query_text(text.clone());
}
if let Some(vec_obj) = vector_arg {
let model = vec_obj
.get("model")
.and_then(Value::as_str)
.ok_or_else(|| anyhow!("vector.model missing"))?
.to_string();
let vals = vec_obj
.get("values")
.and_then(Value::as_array)
.ok_or_else(|| anyhow!("vector.values missing"))?;
let mut vector: Vec<f32> = Vec::with_capacity(vals.len());
for v in vals {
let f = v
.as_f64()
.ok_or_else(|| anyhow!("vector.values element is not a number"))?;
vector.push(f as f32);
}
r = r.vector(model, vector);
} else if let Some(ref text) = text_arg {
// Path A audit fix (2026-04-26): when caller passes `text`
// without an explicit `vector`, try to resolve an embedder
// (env → per-repo config.toml → bundled MiniLM) and embed the
// text in-process. With `--features bundled-embedder`, this
// makes `mnem_retrieve` semantic by default - no Ollama, no
// explicit vector arg, no `MNEM_EMBED_*` setup. Without the
// feature and without env / config, this is a silent no-op
// (current behaviour preserved): the dense lane is skipped
// and the existing graceful-empty-success path handles the
// result.
//
// Failures (provider open error, embed error) are also
// silent: the dense lane is skipped, structured filters
// continue to work. Surfacing the error here would regress
// the no-config-empty-repo case.
#[cfg(feature = "summarize")]
{
if let Some(emb_cfg) = crate::tools::embed::resolve_embed_cfg(repo_path)
&& let Ok(embedder) = mnem_embed_providers::open(&emb_cfg)
&& let Ok(vec) = embedder.embed(text)
{
let model = embedder.model().to_string();
r = r.vector(model, vec);
}
}
}
if let Some(budget) = args.get("token_budget").and_then(Value::as_u64) {
r = r.token_budget(budget.min(u64::from(u32::MAX)) as u32);
}
// Input clamps mirror mnem http: MCP tool args are as untrusted
// as an HTTP body. Without a ceiling, a caller can send
// `limit=18446744073709551615` and trigger whatever the
// downstream BruteForce vector search allocates. See
// `mnem_http::handlers::{MAX_RETRIEVE_LIMIT, MAX_VECTOR_CAP,
// MAX_RERANK_TOP_K}` for the same constants.
if let Some(limit) = args.get("limit").and_then(Value::as_u64) {
let limit = limit as usize;
if limit > MAX_RETRIEVE_LIMIT {
bail!("limit={limit} exceeds max of {MAX_RETRIEVE_LIMIT}");
}
r = r.limit(limit);
}
if let Some(cap) = args.get("vector_cap").and_then(Value::as_u64) {
let cap = cap as usize;
if cap > MAX_VECTOR_CAP {
bail!("vector_cap={cap} exceeds max of {MAX_VECTOR_CAP}");
}
r = r.vector_cap(cap);
}
if let Some(k) = args.get("rerank_top_k").and_then(Value::as_u64) {
let k = k as usize;
if k > MAX_RERANK_TOP_K {
bail!("rerank_top_k={k} exceeds max of {MAX_RERANK_TOP_K}");
}
r = r.rerank_top_k(k);
}
// Experiment E1: community-filter knobs. Accepted at the MCP
// surface for HTTP<->MCP parity, but a live community lookup is
// not wired yet (future work: server-level cache keyed by commit
// CID). When no lookup is installed, passing
// `community_filter=true` is a byte-exact pass-through per the
// retriever's flag-off contract.
let _community_filter = args
.get("community_filter")
.and_then(Value::as_bool)
.unwrap_or(false);
let _community_min_coverage = args
.get("community_min_coverage")
.and_then(Value::as_f64)
.map(|v| v as f32);
if let Some(strategy) = args.get("fusion").and_then(Value::as_str) {
let fusion = match strategy {
"rrf" => mnem_core::retrieve::FusionStrategy::Rrf,
"convex_min_max" => mnem_core::retrieve::FusionStrategy::ConvexMinMax,
other => bail!("fusion must be one of 'convex_min_max' or 'rrf'; got '{other}'"),
};
r = r.fusion(fusion);
}
// Graph-expand block: the only *required* knob is `graph_expand`
// (the global frontier cap). Every other graph_* knob takes the
// retriever default when absent. Keep the same semantics as
// POST /v1/retrieve for MCP<->HTTP parity.
if let Some(max_expand) = args.get("graph_expand").and_then(Value::as_u64) {
let mut cfg = mnem_core::retrieve::GraphExpand {
max_expand: max_expand as usize,
decay: args
.get("graph_decay")
.and_then(Value::as_f64)
.map_or(mnem_core::retrieve::GraphExpand::DEFAULT_DECAY, |d| {
d as f32
}),
etype_filter: args
.get("graph_etype")
.and_then(|v| v.as_array())
.map(|arr| {
arr.iter()
.filter_map(|e| e.as_str().map(ToString::to_string))
.collect()
}),
..Default::default()
};
if let Some(depth) = args.get("graph_depth").and_then(Value::as_u64) {
cfg = cfg.with_depth(depth as usize);
}
if let Some(cap) = args.get("graph_max_per_seed").and_then(Value::as_u64) {
cfg = cfg.with_max_per_seed(cap as usize);
}
// E2: PPR mode dispatch. Parity with POST /v1/retrieve.
if let Some(mode) = args.get("graph_mode").and_then(Value::as_str)
&& mode == "ppr"
{
let damping = args
.get("ppr_damping")
.and_then(Value::as_f64)
.map_or(mnem_core::ppr::DEFAULT_DAMPING, |d| d as f32);
let iter = args
.get("ppr_iter")
.and_then(Value::as_u64)
.map_or(mnem_core::ppr::DEFAULT_MAX_ITER, |n| n as u32);
cfg = cfg.with_ppr(damping, iter, mnem_core::ppr::DEFAULT_EPS);
}
r = r.with_graph_expand(cfg);
}
// audit-2026-04-25 C3-1 (Cycle-3, partial): a fresh repo with no
// `[embed]` config and no `where`/`label`/`vector` filters
// produces `RetrievalEmpty` from the retriever pipeline. The
// error message is accurate but Pass-2 found that MCP clients
// see the bounce and assume the tool is broken. Convert that
// specific `RetrievalEmpty` outcome into an empty-success
// response with a human-readable hint pointing at the
// remediation (`embed.provider` config or `where`/`label`
// filters). All other retrieval failures still propagate as
// tool-level errors.
//
// FOLLOW-UP (deferred from one-pass scope): also wire a default
// mock embedder so a query+limit-only call can return real hits
// on a fresh repo. That is risk-bearing surgery (touches
// `Server` lane init); keeping the conservative no-config
// empty-success here, gated by the existing
// `mnem.v1.retrieve.empty` shape, until the embedder default is
// ready.
let result = match r.execute() {
Ok(r) => r,
Err(e) => {
// Match by message; the retriever surfaces
// `mnem_core::error::RepoError::RetrievalEmpty`
// through anyhow with the canonical text "retrieve:
// no filters or rankers configured.". Detect the
// prefix to avoid swallowing other retrieve errors.
let msg = format!("{e:#}");
if msg.contains("no filters or rankers configured") {
let mut out = String::new();
out.push_str(
"mnem_retrieve: 0 item(s), 0/unlimited tokens, 0 dropped, 0 candidates\n",
);
out.push_str(
" note: no embedder configured and no `where`/`label`/`vector` filter passed; \
returned empty so MCP clients are not bounced on a fresh repo. \
For real hits, either: (a) configure [embed] in <repo>/config.toml or \
`mnem config set embed.provider ollama && mnem config set embed.model nomic-embed-text`, \
or (b) pass `where: {\"key\": \"value\"}` / `label: \"<NodeType>\"` for a pure filter query.\n",
);
return Ok(out);
}
return Err(e.into());
}
};
let mut out = String::new();
out.push_str(&format!(
"mnem_retrieve: {} item(s), {}/{} tokens, {} dropped, {} candidates\n",
result.items.len(),
result.tokens_used,
if result.tokens_budget == u32::MAX {
"unlimited".to_string()
} else {
result.tokens_budget.to_string()
},
result.dropped,
result.candidates_seen,
));
for (i, item) in result.items.iter().enumerate() {
out.push_str(&format!(
" [{i}] score={:.4} tokens={} id={} {}\n",
item.score,
item.tokens,
item.node.id.to_uuid_string(),
item.node.ntype,
));
for line in item.rendered.lines() {
out.push_str(&format!(" {line}\n"));
}
}
Ok(out)
}