ctxmgr

package
v1.166.1 Latest Latest
Warning

This package is not in the latest version of its module.

 Go to latest Published: Jul 20, 2026 License: Apache-2.0 Imports: 23 Imported by: 0

Documentation

Overview

* ChatCLI - Command Line Interface for LLM interaction * Copyright (c) 2024 Edilson Freitas * License: Apache-2.0

* ChatCLI - Knowledge index card (digest) builder. * Copyright (c) 2024 Edilson Freitas * License: Apache-2.0 * * A knowledge context never injects its corpus: attaching one puts only this * digest in the system prompt — a stable, budget-bounded table of contents * that tells the model WHAT the knowledge base covers and HOW to reach it * (passages are auto-retrieved per turn; agent/coder can additionally pull on * demand). A 6MB corpus and a 60MB corpus cost the same handful of tokens per * turn. The output is deterministic for a given context, so it lives in the * cached prompt prefix without busting provider caches.

* ChatCLI - Knowledge-mode ingestion for /context. * Copyright (c) 2024 Edilson Freitas * License: Apache-2.0 * * Knowledge mode turns a flattened documentation corpus (the JSONL emitted by * @docs-flatten, or any directory of docs) into a retrieval-first knowledge * base: the conversation receives only a compact index card, and passages are * pulled on demand. This file owns the ingestion side — parsing the * docs-flatten JSONL schema into the context's file list so every chunk keeps * its provenance (source path, title, repo, commit) instead of arriving as one * opaque multi-megabyte text file.

* ChatCLI - Knowledge-base query surface for the @knowledge tool. * Copyright (c) 2024 Edilson Freitas * License: Apache-2.0 * * PR 1 made knowledge contexts cheap to attach (index card + per-turn push); * this file is the pull side: the manager methods the @knowledge tool uses so * the agent can interrogate an attached corpus on demand — search passages, * read a whole source document, walk the table of contents. Everything is * budget-bounded and works keyless (hybrid retrieval has a BM25 floor).

* ChatCLI - Keyless lexical retrieval (BM25) for knowledge contexts. * Copyright (c) 2024 Edilson Freitas * License: Apache-2.0 * * The embedding-backed retrieval engine needs an API key (Voyage/OpenAI/ * Bedrock); knowledge mode must work without one — the project's keyless-first * rule. This file is that floor: a small pure-Go BM25 index over the same * Segment grain the vector path uses. It is built in memory on demand (a 6MB * corpus tokenizes in well under a second) and combined with cosine scores by * the hybrid retriever when embeddings are available.

* ChatCLI - Keyless BM25 ranking over ad-hoc document lists. * Copyright (c) 2024 Edilson Freitas * License: Apache-2.0 * * The knowledge corpus is not the only thing worth ranking without an API * key: saved sessions (and any future in-memory corpus) need the same * language-neutral scoring. This thin exported wrapper reuses the exact * tokenizer and BM25 scorer the knowledge segments use, so ranking behaves * identically across surfaces instead of each caller growing its own ad-hoc * relevance formula.

* ChatCLI - Command Line Interface for LLM interaction * Copyright (c) 2024 Edilson Freitas * License: Apache-2.0

* ChatCLI - Command Line Interface for LLM interaction * Copyright (c) 2024 Edilson Freitas * License: Apache-2.0

* ChatCLI - Command Line Interface for LLM interaction * Copyright (c) 2024 Edilson Freitas * License: Apache-2.0 * * Semantic retrieval engine for /context. * * Raw whole-file injection blows the window on any non-trivial context. The * engine answers that: it segments a context into passages, embeds them once * (cached on disk per context), and at prompt time returns only the top-k * passages relevant to the current question. Provider-agnostic via the shared * embedding layer; a Null/absent provider disables retrieval and the manager * falls back to the legacy whole-content path with zero regression.

* ChatCLI - Command Line Interface for LLM interaction * Copyright (c) 2024 Edilson Freitas * License: Apache-2.0 * * Passage segmentation for semantic /context retrieval. * * The legacy FileChunk groups WHOLE files into ~30k-token buckets — the right * grain for "inject everything under a token budget", the wrong grain for * retrieval: a 30k-token chunk is itself too large to embed meaningfully or to * return as a focused answer. Segment is the retrieval grain: line-aware windows * of a few hundred tokens with a small overlap so a match never falls in a seam. * Whole files stay verbatim for non-RAG attachments; segments exist only to be * embedded and ranked.

* ChatCLI - Command Line Interface for LLM interaction * Copyright (c) 2024 Edilson Freitas * License: Apache-2.0

* ChatCLI - Command Line Interface for LLM interaction * Copyright (c) 2024 Edilson Freitas * License: Apache-2.0

* ChatCLI - Command Line Interface for LLM interaction * Copyright (c) 2024 Edilson Freitas * License: Apache-2.0

Index

Constants

View Source 
const (
	// Tamanho alvo por chunk (em tokens estimados)
	DefaultChunkTargetTokens = 30000 // ~120KB de texto

	// Tamanho máximo por chunk
	MaxChunkTokens = 50000 // ~200KB de texto

	// Tamanho mínimo para considerar dividir
	MinFilesForChunking = 10
)
View Source 
const (
	MaxContextNameLength = 64
	MaxTotalSizeBytes    = 200 * 1024 * 1024 // 200MB
	MinNameLength        = 3
	MaxDescriptionLength = 500
)
View Source 
const (
	// DefaultRetrievalTopK is the passage count injected when --rag is used
	// without an explicit number. Exported so the CLI flag parser and the engine
	// share one source of truth.
	DefaultRetrievalTopK = 8
)

Variables

This section is empty.

Functions

func BuildKnowledgeDigest added in v1.136.0 

func BuildKnowledgeDigest(fc *FileContext, budget int) string

BuildKnowledgeDigest renders the index card for a knowledge context. budget caps the output in bytes; <=0 takes the default. The model-facing scaffolding is literal English on purpose (prompt text, not UI).

func FormatKnowledgeHits added in v1.136.0 

func FormatKnowledgeHits(query string, hits []KnowledgeHit) string

FormatKnowledgeHits renders search results for the tool transcript: each passage cites its knowledge base, source path and position, with a bounded snippet, so the model can follow up with `get` on the exact document.

func FormatKnowledgeSegmentsBlock added in v1.136.0 

func FormatKnowledgeSegmentsBlock(contextName string, segs []Segment) string

FormatKnowledgeSegmentsBlock renders passages pulled from a knowledge base. Same model-facing English scaffolding as FormatSegmentsBlock, with wording that matches the index-card contract: the corpus is searchable, not attached.

func FormatSegmentsBlock added in v1.133.0 

func FormatSegmentsBlock(contextName, query string, segs []Segment) string

FormatSegmentsBlock renders retrieved passages as a prompt block. The format mirrors formatChunk (literal English/emoji, not i18n — this is model-facing content, and the codebase keeps prompt scaffolding in English on purpose) and annotates each passage with its source file and line range so the model can cite precisely and the user can trace what was injected.

Types

type AttachOptions 

type AttachOptions struct {
	Priority       int
	SelectedChunks []int // Vazio = todos os chunks
	RetrievalTopK  int   // > 0 ativa retrieval semântico (top-K trechos por turno)
}

AttachOptions define opções para anexar contextos

type AttachedContext 

type AttachedContext struct {
	ContextID      string    `json:"context_id"`                // ID do contexto
	AttachedAt     time.Time `json:"attached_at"`               // Quando foi anexado
	Priority       int       `json:"priority"`                  // Prioridade na ordem de mensagens (menor = primeiro)
	SelectedChunks []int     `json:"selected_chunks,omitempty"` // CORREÇÃO: Adicionado campo para chunks selecionados
	// RetrievalTopK > 0 turns this attachment into semantic-retrieval mode:
	// instead of injecting the whole content, only the top-K passages relevant
	// to the current turn are injected (query-driven, so it lives in the
	// volatile prompt zone, never the cached prefix). 0 = legacy whole-content.
	RetrievalTopK int `json:"retrieval_top_k,omitempty"`
}

AttachedContext representa um contexto anexado a uma sessão

type ChunkStrategy 

type ChunkStrategy string

ChunkStrategy define a estratégia de divisão 

const (
	ChunkByDirectory ChunkStrategy = "directory" // Agrupar por diretório
	ChunkByFileType  ChunkStrategy = "filetype"  // Agrupar por tipo de arquivo
	ChunkBySize      ChunkStrategy = "size"      // Dividir por tamanho
	ChunkSmart       ChunkStrategy = "smart"     // Estratégia inteligente híbrida
)

type Chunker 

type Chunker struct {
	// contains filtered or unexported fields
}

Chunker divide arquivos em chunks inteligentes

func NewChunker 

func NewChunker(logger *zap.Logger) *Chunker

NewChunker cria uma nova instância de Chunker

func (*Chunker) DivideIntoChunks 

func (c *Chunker) DivideIntoChunks(files []utils.FileInfo, strategy ChunkStrategy) ([]FileChunk, error)

DivideIntoChunks divide arquivos em chunks usando estratégia inteligente

type ContextFilter 

type ContextFilter struct {
	Tags          []string       `json:"tags"`           // Filtrar por tags
	Mode          ProcessingMode `json:"mode"`           // Filtrar por modo
	MinSize       int64          `json:"min_size"`       // Tamanho mínimo
	MaxSize       int64          `json:"max_size"`       // Tamanho máximo
	CreatedAfter  *time.Time     `json:"created_after"`  // Criado após
	CreatedBefore *time.Time     `json:"created_before"` // Criado antes
	NamePattern   string         `json:"name_pattern"`   // Padrão regex para nome
}

ContextFilter filtra contextos ao listar

type ContextMetrics 

type ContextMetrics struct {
	TotalContexts    int            `json:"total_contexts"`
	AttachedContexts int            `json:"attached_contexts"`
	TotalFiles       int            `json:"total_files"`
	TotalSizeBytes   int64          `json:"total_size_bytes"`
	ContextsByMode   map[string]int `json:"contexts_by_mode"`
	LastUpdated      time.Time      `json:"last_updated"`
	StoragePath      string         `json:"storage_path"`
}

ContextMetrics contém métricas sobre o uso de contextos

type DocHit added in v1.164.0 

type DocHit struct {
	Index int
	Score float64
}

DocHit is one ranked document from RankDocsBM25: the input slice index and its (unnormalized) BM25 score.

func RankDocsBM25 added in v1.164.0 

func RankDocsBM25(docs []string, query string, k int) []DocHit

RankDocsBM25 builds a transient BM25 index over docs and returns up to k hits in descending score order (ties break by ascending index, so results are deterministic). Documents that share no term with the query are absent from the result. The index is throwaway by design — session-sized corpora tokenize in milliseconds, and callers with a persistent corpus should use the knowledge store instead.

type FileChunk 

type FileChunk struct {
	Index       int              `json:"index"`        // Índice do chunk (1-based)
	TotalChunks int              `json:"total_chunks"` // Total de chunks
	Files       []utils.FileInfo `json:"files"`        // Arquivos neste chunk
	Description string           `json:"description"`  // Descrição do chunk
	TotalSize   int64            `json:"total_size"`   // Tamanho total
	EstTokens   int              `json:"est_tokens"`   // Tokens estimados
}

FileChunk representa um chunk de arquivos

type FileContext 

type FileContext struct {
	ID          string            `json:"id"`
	Name        string            `json:"name"`
	Description string            `json:"description"`
	Files       []utils.FileInfo  `json:"files"`
	Mode        ProcessingMode    `json:"mode"`
	TotalSize   int64             `json:"total_size"`
	FileCount   int               `json:"file_count"`
	CreatedAt   time.Time         `json:"created_at"`
	UpdatedAt   time.Time         `json:"updated_at"`
	Tags        []string          `json:"tags"`
	Metadata    map[string]string `json:"metadata"`

	ScanOptions         utils.DirectoryScanOptions `json:"-"`
	ScanOptionsMetadata ScanOptionsMetadata        `json:"scan_options_metadata"`

	Chunks        []FileChunk `json:"chunks,omitempty"`         // Chunks divididos (se modo chunked)
	IsChunked     bool        `json:"is_chunked"`               // Se foi dividido em chunks
	ChunkStrategy string      `json:"chunk_strategy,omitempty"` // Estratégia usada
}

FileContext representa um contexto gerenciado contendo arquivos e metadados

type FormatOptions 

type FormatOptions struct {
	IncludeMetadata  bool   `json:"include_metadata"`  // Incluir metadados no prompt
	IncludeTimestamp bool   `json:"include_timestamp"` // Incluir timestamp
	Compact          bool   `json:"compact"`           // Formato compacto (sem índice)
	Role             string `json:"role"`
}

FormatOptions opções para formatar contexto como prompt

type KnowledgeHit added in v1.136.0 

type KnowledgeHit struct {
	ContextName string
	Seg         Segment
}

KnowledgeHit is one retrieved passage tagged with its knowledge base.

type Manager 

type Manager struct {
	Storage *Storage
	// contains filtered or unexported fields
}

Manager gerencia contextos de forma thread-safe

func NewManager 

func NewManager(logger *zap.Logger) (*Manager, error)

NewManager cria uma nova instância do gerenciador de contextos

func (*Manager) AttachContext 

func (m *Manager) AttachContext(sessionID, contextID string, priority int) error

AttachContext anexa um contexto a uma sessão (não envia à LLM ainda)

func (*Manager) AttachContextWithOptions 

func (m *Manager) AttachContextWithOptions(sessionID, contextID string, opts AttachOptions) error

CORREÇÃO 1: Função refatorada para usar a estrutura de dados correta do Manager. AttachContextWithOptions anexa contexto com opções avançadas

func (*Manager) AttachEmbeddingProvider added in v1.133.0 

func (m *Manager) AttachEmbeddingProvider(provider embedding.Provider)

AttachEmbeddingProvider wires (or rewires) the embedding provider that powers semantic /context retrieval. A Null/absent provider still yields a live engine: Enabled() stays false (so --rag attachments degrade to whole content exactly as before), while knowledge-mode hybrid retrieval keeps its keyless BM25 floor. Safe to call once at startup; provider-agnostic across backends.

func (*Manager) AttachedKnowledge added in v1.136.0 

func (m *Manager) AttachedKnowledge(sessionID string) []*FileContext

AttachedKnowledge returns the knowledge-mode contexts attached to the session, sorted by name for deterministic listings.

func (*Manager) BuildPromptMessages 

func (m *Manager) BuildPromptMessages(sessionID string, opts FormatOptions) ([]models.Message, error)

CORREÇÃO 2: Refatorada para usar a estrutura de dados correta e lidar com chunks selecionados. BuildPromptMessages agora considera chunks selecionados

func (*Manager) BuildRetrievedContextMessages added in v1.133.0 

func (m *Manager) BuildRetrievedContextMessages(ctx context.Context, sessionID, query string) ([]models.Message, error)

BuildRetrievedContextMessages runs per-turn retrieval for every attachment that is query-driven — knowledge contexts (always; hybrid BM25+vectors, no API key required) and --rag attachments (vector-only, needs a provider) — and returns one message per context holding only the passages relevant to query. Returns nil when nothing opted in or the query is empty, so the caller can skip the volatile block entirely.

A failure on a single context is logged and skipped, never fatal: a flaky embedding call must not break the turn. The query embedding happens outside the manager lock because it does network I/O.

func (*Manager) CreateContext 

func (m *Manager) CreateContext(ctx context.Context, name, description string, paths []string, mode ProcessingMode, tags []string, force bool) (*FileContext, error)

CreateContext cria um novo contexto a partir de caminhos de arquivos/diretórios

func (*Manager) DeleteContext 

func (m *Manager) DeleteContext(contextID string) error

DeleteContext remove um contexto permanentemente

func (*Manager) DetachContext 

func (m *Manager) DetachContext(sessionID, contextID string) error

DetachContext remove um contexto anexado de uma sessão

func (*Manager) GetAttachedContexts 

func (m *Manager) GetAttachedContexts(sessionID string) ([]*FileContext, error)

GetAttachedContexts retorna os contextos anexados a uma sessão

func (*Manager) GetContext 

func (m *Manager) GetContext(contextID string) (*FileContext, error)

GetContext retorna um contexto pelo ID

func (*Manager) GetContextByName 

func (m *Manager) GetContextByName(name string) (*FileContext, error)

GetContextByName retorna um contexto pelo nome

func (*Manager) GetMetrics 

func (m *Manager) GetMetrics() *ContextMetrics

GetMetrics retorna métricas sobre os contextos

func (*Manager) GetSessionsForContext added in v1.97.0 

func (m *Manager) GetSessionsForContext(contextID string) []string

GetSessionsForContext returns all session IDs that have the given context attached.

func (*Manager) KnowledgeDigest added in v1.136.1 

func (m *Manager) KnowledgeDigest(fc *FileContext) string

KnowledgeDigest returns fc's index card, memoized per context revision. Prompt assembly calls this every turn; without the memo a 50k-passage corpus would pay an O(corpus) walk and sort on each one.

func (*Manager) KnowledgeDocument added in v1.136.0 

func (m *Manager) KnowledgeDocument(sessionID, kb, source string, offset int) (page string, total int, nextOffset int, err error)

KnowledgeDocument returns one page of a source document (all chunks whose source matches, in corpus order), plus pagination info. offset is a character offset into the assembled document; the next offset is returned when more content remains (0 = done).

func (*Manager) KnowledgeDocumentByName added in v1.163.0 

func (m *Manager) KnowledgeDocumentByName(name, source string, offset int) (page string, total int, nextOffset int, err error)

KnowledgeDocumentByName is the catalog-resolved variant of KnowledgeDocument: it reads from a knowledge base by its stored name, independent of any session attachment. Read-only export surface (MCP resources).

func (*Manager) KnowledgeSearch added in v1.136.0 

func (m *Manager) KnowledgeSearch(ctx context.Context, sessionID, kb, query string, k int) ([]KnowledgeHit, error)

KnowledgeSearch runs hybrid retrieval over the attached knowledge bases (one of them when kb is set) and returns up to k passages per base.

func (*Manager) KnowledgeTOC added in v1.136.0 

func (m *Manager) KnowledgeTOC(sessionID, kb, prefix string) (string, error)

KnowledgeTOC lists the source documents of the attached knowledge bases, optionally filtered by a path prefix. Rendering is model-facing English, consistent with the other prompt scaffolding in this package.

func (*Manager) KnowledgeTOCByName added in v1.163.0 

func (m *Manager) KnowledgeTOCByName(name, prefix string) (string, error)

KnowledgeTOCByName is the catalog-resolved variant of KnowledgeTOC: it lists a knowledge base by its stored name, independent of any session attachment. Read-only export surface (MCP resources).

func (*Manager) ListContexts 

func (m *Manager) ListContexts(filter *ContextFilter) ([]*FileContext, error)

ListContexts lista todos os contextos com filtro opcional

func (*Manager) MergeContexts 

func (m *Manager) MergeContexts(name, description string, contextIDs []string, opts MergeOptions) (*FileContext, error)

MergeContexts mescla múltiplos contextos em um novo

func (*Manager) RenderContext added in v1.163.0 

func (m *Manager) RenderContext(name string) (string, error)

RenderContext renders one context's content by name for read-only export (MCP resources). Knowledge contexts render their index card — the corpus itself is read per document via KnowledgeTOCByName/KnowledgeDocumentByName.

func (*Manager) RetrievalEnabled added in v1.133.0 

func (m *Manager) RetrievalEnabled() bool

RetrievalEnabled reports whether a real embedding provider backs retrieval.

func (*Manager) UpdateContext added in v1.35.0 

func (m *Manager) UpdateContext(ctx context.Context, name string, newPaths []string, newMode ProcessingMode, newTags []string, newDescription string) (*FileContext, error)

UpdateContext atualiza um contexto existente

type MergeOptions 

type MergeOptions struct {
	RemoveDuplicates bool     `json:"remove_duplicates"` // Remove arquivos duplicados
	SortByPath       bool     `json:"sort_by_path"`      // Ordena por caminho
	PreferNewer      bool     `json:"prefer_newer"`      // Prefere versões mais recentes em duplicatas
	Tags             []string `json:"tags"`              // Tags para o contexto mesclado
}

MergeOptions configura como contextos devem ser mesclados

type ProcessingMode 

type ProcessingMode string

ProcessingMode define o modo de processamento de arquivos no contexto 

const (
	ModeFull    ProcessingMode = "full"    // Conteúdo completo
	ModeSummary ProcessingMode = "summary" // Apenas estrutura
	ModeChunked ProcessingMode = "chunked" // Dividido em chunks
	ModeSmart   ProcessingMode = "smart"   // Seleção inteligente
	// ModeKnowledge é retrieval-first: o attach injeta só um index card e os
	// trechos relevantes são recuperados por turno (BM25 keyless + embeddings
	// quando configurados) — corpora de vários MB sem estourar a janela.
	ModeKnowledge ProcessingMode = "knowledge"
)

type Processor 

type Processor struct {
	// contains filtered or unexported fields
}

Processor processa arquivos e diretórios para contextos

func NewProcessor 

func NewProcessor(logger *zap.Logger) *Processor

NewProcessor cria uma nova instância de Processor

func (*Processor) EstimateTokenCount 

func (p *Processor) EstimateTokenCount(files []utils.FileInfo) int

EstimateTokenCount estima o número de tokens em um conjunto de arquivos

func (*Processor) ProcessPaths 

func (p *Processor) ProcessPaths(ctx context.Context, paths []string, mode ProcessingMode) ([]utils.FileInfo, utils.DirectoryScanOptions, error)

ProcessPaths processa múltiplos caminhos baseado no modo

type RetrievalEngine added in v1.133.0 

type RetrievalEngine struct {
	// contains filtered or unexported fields
}

RetrievalEngine builds and queries per-context passage vectors.

func NewRetrievalEngine added in v1.133.0 

func NewRetrievalEngine(provider embedding.Provider, baseDir string, logger *zap.Logger) *RetrievalEngine

NewRetrievalEngine wires an engine over an embedding provider. baseDir is the directory where per-context vector caches live (alongside the context JSON). A Null provider is a valid input: vector paths report Enabled()=false while the lexical (keyless) hybrid path stays fully functional.

func (*RetrievalEngine) DropCache added in v1.133.0 

func (e *RetrievalEngine) DropCache(contextID string)

DropCache removes a context's persisted vector file. Called when a context is deleted or its files change wholesale, so no orphaned cache lingers on disk.

func (*RetrievalEngine) Enabled added in v1.133.0 

func (e *RetrievalEngine) Enabled() bool

Enabled reports whether a real embedding provider backs the engine.

func (*RetrievalEngine) Retrieve added in v1.133.0 

func (e *RetrievalEngine) Retrieve(ctx context.Context, fc *FileContext, query string, k int) ([]Segment, error)

Retrieve returns the top-k passages of fc most relevant to query. Segments, the vector-index handle and its prune run come from the same fingerprint cache the hybrid path uses — previously this path re-segmented the context and re-parsed the whole persisted vector JSON on EVERY query, the exact per-call cost the cache exists to avoid. It embeds only segments not already cached, so repeated calls are cheap and never serve a match against stale text.

func (*RetrievalEngine) RetrieveHybrid added in v1.136.0 

func (e *RetrievalEngine) RetrieveHybrid(ctx context.Context, fc *FileContext, query string, k int) ([]Segment, error)

RetrieveHybrid returns the top-k passages of fc most relevant to query, blending keyless BM25 with cosine similarity when an embedding provider is configured. This is the knowledge-mode path: unlike Retrieve it never requires an API key — without a provider it degrades to lexical-only, and a failing embedding call degrades the same way instead of breaking the turn.

Scalability contract: BM25 does RECALL over the whole corpus (in-memory, fingerprint-cached); embeddings only RERANK the candidate pool. Per-query embedding cost is bounded by hybridMaxPool regardless of corpus size — a 60MB corpus is never embedded wholesale, and the vector cache only ever holds passages that some query actually surfaced.

type ScanOptionsMetadata 

type ScanOptionsMetadata struct {
	MaxTotalSize      int64    `json:"max_total_size"`
	MaxFilesToProcess int      `json:"max_files_to_process"`
	Extensions        []string `json:"extensions"`
	ExcludeDirs       []string `json:"exclude_dirs"`
	ExcludePatterns   []string `json:"exclude_patterns"`
	IncludeHidden     bool     `json:"include_hidden"`
}

ScanOptionsMetadata contém versão serializável das opções de scan

type Segment added in v1.133.0 

type Segment struct {
	ID        string // stable content hash — the vector-index key
	FilePath  string
	FileType  string
	StartLine int // 1-based, inclusive
	EndLine   int // 1-based, inclusive
	Content   string
}

Segment is one retrievable passage of a file.

func SegmentFiles added in v1.133.0 

func SegmentFiles(files []utils.FileInfo, opts SegmentOptions) []Segment

SegmentFiles splits every file into overlapping, line-aware passages. The output order is deterministic (file order, then top-to-bottom), and segment ids are content hashes so re-segmenting unchanged files yields identical ids — which lets the vector index skip re-embedding work that hasn't changed.

type SegmentOptions added in v1.133.0 

type SegmentOptions struct {
	MaxChars     int // soft cap per segment (~4 chars/token); default 1200 ≈ 300 tokens
	OverlapLines int // lines replayed at the start of the next segment; default 2
}

SegmentOptions tunes how files are split into passages.

type Storage 

type Storage struct {
	// contains filtered or unexported fields
}

Storage gerencia a persistência de contextos em disco

func NewStorage 

func NewStorage(logger *zap.Logger) (*Storage, error)

NewStorage cria uma nova instância de Storage

func (*Storage) DeleteContext 

func (s *Storage) DeleteContext(contextID string) error

DeleteContext deleta um contexto do disco

func (*Storage) ExportContext 

func (s *Storage) ExportContext(ctx *FileContext, targetPath string) error

ExportContext exporta um contexto para um arquivo específico

func (*Storage) GetStoragePath 

func (s *Storage) GetStoragePath() string

GetStoragePath retorna o caminho base de armazenamento

func (*Storage) ImportContext 

func (s *Storage) ImportContext(sourcePath string) (*FileContext, error)

ImportContext importa um contexto de um arquivo

func (*Storage) LoadAllContexts 

func (s *Storage) LoadAllContexts() ([]*FileContext, error)

LoadAllContexts carrega todos os contextos do disco

func (*Storage) LoadContext 

func (s *Storage) LoadContext(contextID string) (*FileContext, error)

LoadContext carrega um contexto do disco

func (*Storage) SaveContext 

func (s *Storage) SaveContext(ctx *FileContext) error

SaveContext salva um contexto em disco

type ValidationResult 

type ValidationResult struct {
	Valid    bool     `json:"valid"`
	Errors   []string `json:"errors"`
	Warnings []string `json:"warnings"`
}

ValidationResult resultado da validação de um contexto

type Validator 

type Validator struct {
	// contains filtered or unexported fields
}

Validator valida contextos e suas operações

func NewValidator 

func NewValidator(logger *zap.Logger) *Validator

NewValidator cria uma nova instância de Validator

func (*Validator) ValidateContext 

func (v *Validator) ValidateContext(ctx *FileContext) *ValidationResult

ValidateContext valida um contexto completo

func (*Validator) ValidateDescription 

func (v *Validator) ValidateDescription(description string) error

ValidateDescription valida a descrição de um contexto

func (*Validator) ValidateMode 

func (v *Validator) ValidateMode(mode ProcessingMode) error

ValidateMode valida o modo de processamento

func (*Validator) ValidateName 

func (v *Validator) ValidateName(name string) error

ValidateName valida o nome de um contexto

func (*Validator) ValidatePriority 

func (v *Validator) ValidatePriority(priority int) error

ValidatePriority valida a prioridade de anexação

func (*Validator) ValidateTags 

func (v *Validator) ValidateTags(tags []string) error

ValidateTags valida as tags de um contexto

func (*Validator) ValidateTotalSize 

func (v *Validator) ValidateTotalSize(size int64) error

ValidateTotalSize valida o tamanho total de arquivos

Source Files

View all Source files

Jump to

Keyboard shortcuts

? : This menu
/ : Search site
f or F : Jump to
y or Y : Canonical URL
go.dev uses cookies from Google to deliver and enhance the quality of its services and to analyze traffic. Learn more.