Documentation
¶
Index ¶
Constants ¶
const ( // SupersessionPenaltyFactor is the multiplicative penalty for memories that have // been superseded by another memory present in the same result set. SupersessionPenaltyFactor = 0.3 // ConflictPenaltyFactor is the multiplicative penalty for memories with an // active (unresolved) conflict status. ConflictPenaltyFactor = 0.8 )
Variables ¶
var TypePriority = map[models.MemoryType]float64{ models.MemoryTypeRule: 1.5, models.MemoryTypeProcedure: 1.3, models.MemoryTypeFact: 1.0, models.MemoryTypeEpisode: 0.8, models.MemoryTypePreference: 0.7, }
TypePriority maps memory types to their raw priority multipliers (before normalization).
Functions ¶
func FormatWithConflictAnnotations ¶ added in v0.4.0
func FormatWithConflictAnnotations(results []models.RecallResult, budget int) string
FormatWithConflictAnnotations formats recall results with inline conflict annotations. Memories in an active conflict group are annotated with the short IDs of conflicting peers.
Types ¶
type Reasoner ¶ added in v0.4.0
type Reasoner struct {
// contains filtered or unexported fields
}
Reasoner uses Claude to re-rank recall results by genuine relevance to the query. It addresses the "similarity ≠ relevance" problem: vector similarity finds related content, but Claude can reason about which memories are truly useful.
On any API failure the Reasoner degrades gracefully and returns results in their original order so the caller always gets a usable response.
func NewReasoner ¶ added in v0.4.0
NewReasoner creates a Reasoner backed by the Anthropic Claude API.
func (*Reasoner) ReRank ¶ added in v0.4.0
func (r *Reasoner) ReRank(ctx context.Context, query string, results []models.RecallResult, maxCandidates int) ([]models.RecallResult, error)
ReRank passes the top maxCandidates results to Claude and returns them reordered by relevance to query. Results beyond maxCandidates are appended unchanged after the re-ranked set.
If maxCandidates <= 0, defaultReasonerCandidates is used. On any error (API failure, unexpected response) the original order is returned.
type Recaller ¶
type Recaller struct {
// contains filtered or unexported fields
}
Recaller performs multi-factor ranking of search results.
func NewRecaller ¶
NewRecaller creates a new recaller with the given weights. If the weights are invalid, a warning is logged and defaults are used.
func (*Recaller) Rank ¶
func (r *Recaller) Rank(results []models.SearchResult, project string, query string) []models.RecallResult
Rank re-ranks search results using multi-factor scoring. All component scores are normalized to [0,1] before weighting.
func (*Recaller) ShouldRerank ¶ added in v0.4.0
func (r *Recaller) ShouldRerank(results []models.RecallResult, threshold float64) bool
ShouldRerank returns true when the top-4 results are close enough in score that Claude re-ranking may improve ordering. Returns false when:
- threshold is <= 0 (feature disabled)
- fewer than 4 results are provided
- the spread between results[0] and results[3] exceeds the threshold
type Weights ¶
type Weights struct {
Similarity float64 `json:"similarity" mapstructure:"similarity"`
Recency float64 `json:"recency" mapstructure:"recency"`
Frequency float64 `json:"frequency" mapstructure:"frequency"`
TypeBoost float64 `json:"type_boost" mapstructure:"type_boost"`
ScopeBoost float64 `json:"scope_boost" mapstructure:"scope_boost"`
Confidence float64 `json:"confidence" mapstructure:"confidence"`
Reinforcement float64 `json:"reinforcement" mapstructure:"reinforcement"`
TagAffinity float64 `json:"tag_affinity" mapstructure:"tag_affinity"`
}
Weights controls the relative importance of each ranking factor.
func DefaultWeights ¶
func DefaultWeights() Weights
DefaultWeights returns sensible default ranking weights.
Source Files