zeph_memory/semantic/
corrections.rs

1// SPDX-FileCopyrightText: 2026 Andrei G <bug-ops>
2// SPDX-License-Identifier: MIT OR Apache-2.0
3
4use zeph_llm::provider::LlmProvider as _;
5
6use crate::error::MemoryError;
7
8use super::{CORRECTIONS_COLLECTION, SemanticMemory};
9
10impl SemanticMemory {
11    /// Store an embedding for a user correction in the vector store.
12    ///
13    /// Silently skips if no vector store is configured or embeddings are unsupported.
14    ///
15    /// # Errors
16    ///
17    /// Returns an error if embedding generation or vector store write fails.
18    pub async fn store_correction_embedding(
19        &self,
20        correction_id: i64,
21        correction_text: &str,
22    ) -> Result<(), MemoryError> {
23        let Some(ref store) = self.qdrant else {
24            return Ok(());
25        };
26        if !self.provider.supports_embeddings() {
27            return Ok(());
28        }
29        let embedding = self
30            .provider
31            .embed(correction_text)
32            .await
33            .map_err(|e| MemoryError::Other(e.to_string()))?;
34        let vector_size = u64::try_from(embedding.len()).unwrap_or(896);
35        store
36            .ensure_named_collection(CORRECTIONS_COLLECTION, vector_size)
37            .await?;
38        let payload = serde_json::json!({ "correction_id": correction_id });
39        store
40            .store_to_collection(CORRECTIONS_COLLECTION, payload, embedding)
41            .await?;
42        Ok(())
43    }
44
45    /// Retrieve corrections semantically similar to `query`.
46    ///
47    /// Returns up to `limit` corrections scoring above `min_score`.
48    /// Returns an empty vec if no vector store is configured.
49    ///
50    /// # Errors
51    ///
52    /// Returns an error if embedding generation or vector search fails.
53    pub async fn retrieve_similar_corrections(
54        &self,
55        query: &str,
56        limit: usize,
57        min_score: f32,
58    ) -> Result<Vec<crate::sqlite::corrections::UserCorrectionRow>, MemoryError> {
59        let Some(ref store) = self.qdrant else {
60            tracing::debug!("corrections: skipped, no vector store");
61            return Ok(vec![]);
62        };
63        if !self.provider.supports_embeddings() {
64            tracing::debug!("corrections: skipped, no embedding support");
65            return Ok(vec![]);
66        }
67        let embedding = self
68            .provider
69            .embed(query)
70            .await
71            .map_err(|e| MemoryError::Other(e.to_string()))?;
72        let vector_size = u64::try_from(embedding.len()).unwrap_or(896);
73        store
74            .ensure_named_collection(CORRECTIONS_COLLECTION, vector_size)
75            .await?;
76        let scored = store
77            .search_collection(CORRECTIONS_COLLECTION, &embedding, limit, None)
78            .await
79            .unwrap_or_default();
80
81        tracing::debug!(
82            candidates = scored.len(),
83            min_score = %min_score,
84            limit,
85            "corrections: search complete"
86        );
87
88        let mut results = Vec::new();
89        for point in scored {
90            if point.score < min_score {
91                continue;
92            }
93            if let Some(id_val) = point.payload.get("correction_id")
94                && let Some(id) = id_val.as_i64()
95            {
96                let rows = self.sqlite.load_corrections_for_id(id).await?;
97                results.extend(rows);
98            }
99        }
100
101        tracing::debug!(
102            retained = results.len(),
103            "corrections: after min_score filter"
104        );
105
106        Ok(results)
107    }
108}
zeph_memory/semantic/corrections.rs

zeph_memory/semantic/
corrections.rs
