@startuml Contextual Embeddings System
!include shared_interfaces.puml
package "Contextual Embedding System" {
class ContextualEmbeddingEngine {
- embedder: VoyageEmbedder
- contextBuilder: ContextBuilder
- cache: EmbeddingCache
+ embedChunk(chunk: DocumentChunk): float[]
+ embedWithContext(content, context): float[]
+ batchEmbed(chunks: DocumentChunk[]): float[][]
}
class ContextBuilder {
- summaryGenerator: TextSummarizer
+ buildContext(chunk: DocumentChunk): ContextualText
+ generateSummary(text: string, maxLen: int): string
+ extractSectionPath(chunk): string[]
+ mergeContextWindows(before, content, after): string
}
class ContextualText {
+ documentTitle: string
+ sectionPath: string[]
+ contextBefore: string
+ mainContent: string
+ contextAfter: string
+ metadata: dict
+ toString(): string
}
class AdaptiveChunker {
- strategies: Map<string, ChunkStrategy>
+ chunkDocument(doc: Document): DocumentChunk[]
+ selectStrategy(doc): ChunkStrategy
+ applyOverlap(chunks): DocumentChunk[]
+ validateChunks(chunks): bool
}
abstract class ChunkStrategy {
# targetSize: int
# overlapSize: int
+ {abstract} chunk(content): Chunk[]
+ {abstract} findBoundaries(content): int[]
# preserveContext(chunk): Chunk
}
class SimpleChunkStrategy {
+ chunk(content): Chunk[]
+ findBoundaries(content): int[]
- splitBySize(content): string[]
}
class HierarchicalChunkStrategy {
- sectionParser: SectionParser
+ chunk(content): Chunk[]
+ findBoundaries(content): int[]
- splitBySections(content): Section[]
- balanceSectionSizes(sections): Chunk[]
}
class SemanticChunkStrategy {
- topicModeler: TopicModeler
- sentenceEncoder: SentenceTransformer
+ chunk(content): Chunk[]
+ findBoundaries(content): int[]
- clusterBySimilarity(sentences): Cluster[]
- optimizeClusterSizes(clusters): Chunk[]
}
class EmbeddingCache {
- storage: RedisCache
- ttl: int
+ get(key: string): float[]
+ set(key: string, embedding: float[])
+ batch_get(keys: string[]): Map<string, float[]>
+ invalidate(pattern: string)
}
}
package "Hybrid Search System" {
class HybridSearchEngine {
- bm25Index: BM25Index
- vectorStore: QdrantClient
- reranker: CrossEncoderReranker
+ search(query: string, options): SearchResult[]
- searchBM25(query): BM25Result[]
- searchVector(query): VectorResult[]
- mergeResults(bm25, vector): CombinedResult[]
}
class BM25Index {
- sqliteStore: SQLiteStore
+ index(chunks: DocumentChunk[])
+ search(query: string, limit: int): BM25Result[]
+ updateIndex(chunk: DocumentChunk)
+ getTermFrequencies(query): Map<string, float>
}
class CrossEncoderReranker {
- model: CrossEncoderModel
- batchSize: int
+ rerank(query: string, candidates: Result[]): Result[]
+ scoreCandidate(query, candidate): float
+ batchScore(query, candidates): float[]
}
class SearchResult {
+ chunkId: string
+ content: string
+ contextualContent: string
+ score: float
+ bm25Score: float
+ vectorScore: float
+ rerankScore: float
+ metadata: ResultMetadata
}
}
' Relationships
ContextualEmbeddingEngine --> ContextBuilder : uses
ContextualEmbeddingEngine --> EmbeddingCache : caches in
ContextBuilder --> ContextualText : creates
AdaptiveChunker --> ChunkStrategy : selects
SimpleChunkStrategy --|> ChunkStrategy
HierarchicalChunkStrategy --|> ChunkStrategy
SemanticChunkStrategy --|> ChunkStrategy
HybridSearchEngine --> BM25Index : searches
HybridSearchEngine --> ContextualEmbeddingEngine : uses
HybridSearchEngine --> CrossEncoderReranker : reranks with
' External dependencies
ContextualEmbeddingEngine ..> VoyageAI : embeds with
BM25Index ..> SQLiteStore : queries
HybridSearchEngine ..> QdrantClient : searches vectors
@enduml
