argus

package module
v0.1.0 Latest Latest
Warning

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

 Go to latest Published: Aug 24, 2026 License: MIT Imports: 29 Imported by: 0

README

Argus

Go Reference

Argus is a Go library that indexes source code into a knowledge graph and answers natural language questions about it. It combines semantic search, BM25, call-graph traversal, and cross-encoder reranking to surface relevant symbols with citations.

// 1. Create an engine
engine, _ := argus.New(argus.Config{
    // Required
    FalkorDBAddr: "localhost:6382",
    TEIURL:       "http://localhost:8085",

    // Optional: cross-encoder reranker for sharper relevance scoring
    RerankURL: "http://localhost:8086",

    // Optional: LLM for synthesized answers and component enrichment
    LLMBaseURL: "http://localhost:11434",
    LLMModel:   "llama3.1:8b",
    LLMAPIKey:  "", // set if your LLM requires auth
})
defer engine.Close()

// 2. Index a codebase
result, _ := engine.Indexer.Index(ctx, argus.IndexRequest{Path: "./my-project"})

// 3. Ask a question
answer, _ := engine.Ask(ctx, "How does error handling work?")
fmt.Println(answer.Answer) // synthesized answer with file:line citations

Features

  • Multi-language parsing — Go, Rust, JavaScript/TypeScript, Python via tree-sitter
  • Hybrid retrieval — Semantic + BM25 + call-graph BFS, fused with RRF
  • Cross-encoder reranking — TEI-powered joint relevance scoring
  • Component detection — Auto-groups symbols into architectural components with dependency edges
  • Component-scoped queries — Narrow search to a single component and its dependencies
  • Call graph traversal — Follow caller/callee chains from any function
  • Optional LLM synthesis — Ollama/OpenAI integration for natural language answers

Installation

# CLI (interactive TUI)
go install github.com/akhilparakka/argus/cmd/argus@latest

# Library (import into your own code)
go get github.com/akhilparakka/argus

Quick Start

1. Start FalkorDB and TEI

Argus needs two services: FalkorDB (graph database) and TEI (text embeddings inference).

# Option A: Docker Compose (local)
docker compose up -d

# Option B: Your own instances
# export FALKORDB_ADDR="falkordb.mydomain.com:6379"
# export TEI_URL="https://tei.mydomain.com"
2. Create an engine
engine, err := argus.New(argus.Config{
    // Required
    FalkorDBAddr: "localhost:6382",
    TEIURL:       "http://localhost:8085",

    // Optional: LLM for synthesized answers (Ollama, OpenAI, etc.)
    LLMBaseURL: "http://localhost:11434",
    LLMModel:   "llama3.1:8b",
    LLMAPIKey:  "", // set if your LLM requires auth

    // Optional: reranker for sharper relevance scoring
    RerankURL: "http://localhost:8086",
})
if err != nil {
    log.Fatal(err)
}
defer engine.Close()
3. Index a codebase
result, err := engine.Indexer.Index(ctx, argus.IndexRequest{
    Path: "/path/to/codebase",
})
fmt.Printf("Indexed %d files, %d symbols\n", result.FilesProcessed, result.NodesCreated)
4. Ask questions
// With LLM configured: returns a synthesized natural language answer
result, _ := engine.Ask(ctx, "How does the agent loop work?")
fmt.Println(result.Answer)

// Without LLM: returns raw formatted context
result, _ := engine.Ask(ctx, "Show me error handling code")
fmt.Println(result.Context) // ranked symbols with file:line citations

// Scoped to a single component
result, _ := engine.Ask(ctx, "What does the embedding service do?",
    &argus.ComponentScope{Name: "Embedding Service"},
)

Configuration

Field Required Default Description
FalkorDBAddr ✅ — FalkorDB host:port (Redis protocol)
TEIURL ✅ — TEI (text-embeddings-inference) endpoint
GraphName "argus" FalkorDB graph name
RerankURL — Cross-encoder reranker endpoint
RerankToken — Bearer token for authenticated reranker
LLMBaseURL — LLM API base URL (Ollama, OpenAI, etc.)
LLMModel — Model name for LLM synthesis
LLMAPIKey — API key for authenticated LLM endpoints
Workers 8 Concurrent indexer workers
SemanticLimit 20 Semantic search top-K
BM25Limit 20 BM25 search top-K
CallGraphDepth 2 Call-graph BFS expansion depth
ResultCap 30 Final result cap before rerank

Low-Level API

For custom pipelines, use the client directly:

// Search APIs
semantic, _ := client.SearchCodeSemantic(ctx, embedding, "", nil, 20)
bm25, _ := client.SearchCodeBM25(ctx, "error handling", "", nil, 20)
cg, _ := client.GetCallGraph(ctx, "pkg.(*Server).Handle", "forward", 3, 0)

// Component APIs
comps, _ := client.GetComponentTree(ctx)

CLI

A terminal UI is included for interactive exploration.

# Setup (one-time)
argus init

# Run
argus

Features: component picker (Tab), source expansion (Ctrl+E), text copy (Ctrl+Y), markdown rendering.

Graph Schema

Nodes (:Code): functions, methods, structs, interfaces, files, packages, components.

Edges: DEFINES, CALLS, IMPLEMENTS, IMPORTS, CONTAINS, PART_OF, HAS_METHOD.

Indexes: vector index on code_embedding, full-text index on doc_comment.

Development

git clone https://github.com/akhilparakka/argus
cd argus
docker compose up -d
go test ./...

License

MIT

Documentation

Index

Constants

View Source 
const CodeChatSystemPrompt = `` /* 418-byte string literal not displayed */

CodeChatSystemPrompt instructs the synthesis LLM to answer strictly from retrieved code context. Exported so downstream chat implementations (TUI, Terminalator) share one prompt instead of drifting copies.

View Source 
const (
	DefaultRRFK = 60
)

Retrieval tuning constants.

View Source 
const DefaultVectorDimension = 768

DefaultVectorDimension is the fallback embedding dimension used when the caller does not specify one. It matches the BAAI/bge-base-en-v1.5 TEI model.

Variables

View Source 
var GoPlugin = &LanguagePlugin{
	Name:         LangGo,
	Extensions:   []string{".go"},
	NewParser:    func() (*Parser, error) { return NewParser(LangGo) },
	NewExtractor: func() (*Extractor, error) { return NewExtractor(LangGo) },
	Qualifier:    &GoQualifier{},
}
View Source 
var JavaScriptPlugin = &LanguagePlugin{
	Name:         LangJavaScript,
	Extensions:   []string{".js", ".jsx", ".mjs"},
	NewParser:    func() (*Parser, error) { return NewParser(LangJavaScript) },
	NewExtractor: func() (*Extractor, error) { return NewExtractor(LangJavaScript) },
	Qualifier:    &JSQualifier{},
}
View Source 
var PythonPlugin = &LanguagePlugin{
	Name:         LangPython,
	Extensions:   []string{".py", ".pyw"},
	NewParser:    func() (*Parser, error) { return NewParser(LangPython) },
	NewExtractor: func() (*Extractor, error) { return NewExtractor(LangPython) },
	Qualifier:    &PythonQualifier{},
}
View Source 
var RustPlugin = &LanguagePlugin{
	Name:         LangRust,
	Extensions:   []string{".rs"},
	NewParser:    func() (*Parser, error) { return NewParser(LangRust) },
	NewExtractor: func() (*Extractor, error) { return NewExtractor(LangRust) },
	Qualifier:    &RustQualifier{},
}

Functions

func EnrichComponents 

func EnrichComponents(
	ctx context.Context,
	p LLMProvider,
	components []CodeNode,
	symbolSummaries map[string][]string,
) (map[string]ComponentEnrichment, error)

EnrichComponents calls the LLM once with a batched prompt to generate human-readable names, roles, and descriptions for each component. Returns a map from directory path → enrichment. On failure the caller should keep the path-based names.

func FetchTEIDimension 

func FetchTEIDimension(baseURL string) (int, error)

FetchTEIDimension discovers the model's embedding dimension by sending a single probe embedding request and measuring the returned vector length.

func PingFalkorDB 

func PingFalkorDB(addr string) error

PingFalkorDB verifies that a FalkorDB instance is reachable at addr. It returns a clear error for connection failures vs non-FalkorDB servers.

func PingOllama 

func PingOllama(baseURL, apiKey string) error

PingOllama verifies the Ollama endpoint is reachable. It checks the /api/tags endpoint (local) or simply that the base URL responds. Model availability is checked at runtime when the user actually sends a chat request.

func PingTEI 

func PingTEI(baseURL string) error

PingTEI sends a lightweight request to verify the TEI endpoint is alive.

func RegisterLanguage 

func RegisterLanguage(p *LanguagePlugin)

RegisterLanguage registers a language plugin for all its file extensions.

func RegisteredExtensions 

func RegisteredExtensions() []string

RegisteredExtensions returns all registered file extensions.

Types

type BuilderOption 

type BuilderOption func(*CodeContextBuilder)

BuilderOption configures optional CodeContextBuilder behavior.

func WithBuilderLogger 

func WithBuilderLogger(l *slog.Logger) BuilderOption

WithBuilderLogger sets a logger for retrieval diagnostics (search failures, degenerate embeddings, rerank problems). If unset, the builder runs silent.

type CallGraph 

type CallGraph struct {
	Root      string     `json:"root"`      // qualified_name
	Direction string     `json:"direction"` // forward | reverse
	Depth     int        `json:"depth"`
	Nodes     []CodeNode `json:"nodes"`
	Edges     []CodeEdge `json:"edges"`
}

CallGraph represents a function's call hierarchy.

type CallSite 

type CallSite struct {
	CalleeName   string // function/method name being called: "HandleLogin", "GetUser"
	Object       string // receiver/object for method calls: "db", "s", "" for bare calls
	Line         int    // line where the call happens (1-based)
	IsMethodCall bool   // true = obj.Method(), false = Function()
}

CallSite represents a function or method call found in source code.

type CodeChatResult 

type CodeChatResult struct {
	Query   string
	Context string // raw context sent to the LLM
	Answer  string // LLM-synthesized answer (or context if no LLM)
	Sources []CodeNode
	Error   string

	// ScopeResolved is false when a ComponentScope was requested but matched
	// no component in the graph (stale name after re-enrichment, typo, or
	// removed directory). Results are then empty for that reason — UIs should
	// refresh their component list instead of showing "not enough context".
	ScopeResolved bool
}

CodeChatResult is the output of a single code-chat query.

type CodeContext 

type CodeContext struct {
	Nodes   []CodeNode
	Query   string
	Content string // formatted context string for LLM consumption

	// ScopeResolved reports whether a requested ComponentScope actually
	// matched a component in the graph. False + non-nil scope means the
	// caller used a stale/unknown name (e.g. after LLM enrichment renamed
	// it) — results are empty for that reason, NOT because the question
	// was hard. Surfaces in CodeChatResult so UIs can self-heal.
	ScopeResolved bool
}

CodeContext is the result of a code retrieval pipeline.

type CodeContextBuilder 

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

CodeContextBuilder retrieves relevant code symbols for a natural language query. Pipeline: embed query -> semantic + BM25 (concurrent) -> callgraph BFS -> RRF fusion -> cross-encoder rerank -> format.

func NewCodeContextBuilder 

func NewCodeContextBuilder(client *FalkorClient, embedder Embedder, rerankURL, rerankToken string, opts ...BuilderOption) *CodeContextBuilder

NewCodeContextBuilder creates a builder backed by the given FalkorDB client and embedder. rerankURL and rerankToken are optional; if empty, reranking is skipped.

func (*CodeContextBuilder) Build 

func (b *CodeContextBuilder) Build(ctx context.Context, params CodeContextBuilderParams) (CodeContext, error)

Build runs the code retrieval pipeline.

type CodeContextBuilderParams 

type CodeContextBuilderParams struct {
	Query string
	Scope *ComponentScope // nil = unscoped (search all symbols)
}

CodeContextBuilderParams holds the user query and workspace information.

type CodeEdge 

type CodeEdge struct {
	SourceUID string `json:"source_uid"`
	TargetUID string `json:"target_uid"`
	Type      string `json:"type"`           // DEFINES | CALLS | IMPLEMENTS | IMPORTS | CONTAINS | PART_OF | HAS_METHOD
	Line      int    `json:"line,omitempty"` // for CALLS: line where call happens
}

CodeEdge represents a relationship between two code entities.

type CodeNode 

type CodeNode struct {
	UID           string    `json:"uid,omitempty"`
	Name          string    `json:"name"`
	QualifiedName string    `json:"qualified_name"`
	Kind          string    `json:"kind"` // function | method | struct | interface | class | trait | constant | variable | file | package | component
	FilePath      string    `json:"file_path"`
	StartLine     int       `json:"start_line"`
	EndLine       int       `json:"end_line"`
	Signature     string    `json:"signature"`
	Receiver      string    `json:"receiver,omitempty"` // methods only
	Language      string    `json:"language"`
	IsExported    bool      `json:"is_exported"`
	DocComment    string    `json:"doc_comment"`
	Role          string    `json:"role,omitempty"`
	FileHash      string    `json:"file_hash,omitempty"` // file nodes only
	IndexedAt     int64     `json:"indexed_at"`
	Embedding     []float32 `json:"code_embedding,omitempty"`
	Score         float64   `json:"score,omitempty"` // search relevance score (not stored in DB)
}

CodeNode represents a code entity in the knowledge graph.

func DetectComponents 

func DetectComponents(nodes []CodeNode, rootPath string) []CodeNode

DetectComponents groups symbols by directory into component CodeNodes. Each unique directory becomes a component. The name is set to the relative directory path; if an LLM provider is configured the indexer overwrites it with a human-readable name via EnrichComponents.

type ComponentEnrichment 

type ComponentEnrichment struct {
	Dir         string `json:"dir"`
	Name        string `json:"name"`
	Role        string `json:"role"`
	Description string `json:"description"`
}

ComponentEnrichment holds the LLM-generated metadata for a component.

type ComponentNode 

type ComponentNode struct {
	Name        string   `json:"name"`
	Role        string   `json:"role"`
	Description string   `json:"description"`
	Color       string   `json:"color"`
	FilePath    string   `json:"file_path"`
	FileCount   int      `json:"file_count"`
	Symbols     []string `json:"symbols"`
	Downstream  []string `json:"downstream"`
	Upstream    []string `json:"upstream"`
}

ComponentNode represents a detected component.

type ComponentScope 

type ComponentScope struct {
	Name        string // component name to focus on
	IncludeDeps bool   // also search symbols from dependency components
	MaxDeps     int    // cap on dependency components (0 = no cap)
}

ComponentScope restricts retrieval to symbols belonging to a specific component. When passed to Engine.Ask, the retrieval pipeline filters semantic and BM25 results to only include symbols with PART_OF edges to the named component (and optionally its direct dependencies).

type Config 

type Config struct {
	// Service URLs (required)
	FalkorDBAddr string // e.g. "localhost:6382"
	TEIURL       string // e.g. "http://localhost:8085"

	// Graph name (default: "argus")
	GraphName string

	// Reranker (optional)
	RerankURL   string
	RerankToken string

	// LLM (optional — only for chat synthesis and component enrichment)
	LLMBaseURL string
	LLMModel   string
	LLMAPIKey  string

	// Tuning
	Workers        int // Indexer workers (default 8)
	RRFK           int // RRF k parameter (default 60)
	SemanticLimit  int // Semantic search result limit (default 20)
	BM25Limit      int // BM25 search result limit (default 20)
	CallGraphDepth int // BFS expansion depth (default 2)
	ResultCap      int // Final result cap (default 30)

	// Optional logger for indexer and retrieval diagnostics.
	// If nil, both run silent.
	Logger *slog.Logger
}

Config holds all service configuration and tuning parameters for the Argus engine.

type Embedder 

type Embedder interface {
	Embed(ctx context.Context, texts []string) ([][]float32, error)
	Dimension() int
}

Embedder turns text into dense vectors for semantic search.

type Engine 

type Engine struct {
	Client   *FalkorClient
	Embedder Embedder
	Indexer  *Indexer

	Config Config
	// contains filtered or unexported fields
}

Engine is the fully wired Argus engine with indexing, retrieval, and chat.

func New 

func New(cfg Config) (*Engine, error)

New creates a fully wired Engine from the given Config.

Construction order matters: the TEI embedder is created first so its auto-detected embedding dimension can size the FalkorDB vector index. Creating the client first would hardcode the dimension and break every embedding write against non-768 TEI models (e.g. bge-m3).

func (*Engine) Ask 

func (e *Engine) Ask(ctx context.Context, query string, scope ...*ComponentScope) (CodeChatResult, error)

Ask runs the code retrieval pipeline and optionally calls the LLM to synthesize an answer. If no LLM is configured, it returns the raw formatted context as the answer. An optional ComponentScope restricts retrieval to symbols belonging to the named component.

func (*Engine) Close 

func (e *Engine) Close() error

Close cleans up all resources held by the Engine.

type ErrEmbeddingFailed 

type ErrEmbeddingFailed struct {
	Cause error
}

ErrEmbeddingFailed is returned when text-to-vector conversion fails.

func (ErrEmbeddingFailed) Error 

func (e ErrEmbeddingFailed) Error() string

func (ErrEmbeddingFailed) Unwrap 

func (e ErrEmbeddingFailed) Unwrap() error

type ErrFalkorDBUnavailable 

type ErrFalkorDBUnavailable struct {
	Addr string
	Err  error
}

ErrFalkorDBUnavailable is returned when FalkorDB cannot be reached.

func (ErrFalkorDBUnavailable) Error 

func (e ErrFalkorDBUnavailable) Error() string

func (ErrFalkorDBUnavailable) Unwrap 

func (e ErrFalkorDBUnavailable) Unwrap() error

type ErrInvalidQuery 

type ErrInvalidQuery struct {
	Query string
}

ErrInvalidQuery is returned when the query is empty or too short.

func (ErrInvalidQuery) Error 

func (e ErrInvalidQuery) Error() string

type ErrLLMUnavailable 

type ErrLLMUnavailable struct {
	BaseURL string
	Err     error
}

ErrLLMUnavailable is returned when the LLM provider cannot be reached.

func (ErrLLMUnavailable) Error 

func (e ErrLLMUnavailable) Error() string

func (ErrLLMUnavailable) Unwrap 

func (e ErrLLMUnavailable) Unwrap() error

type ErrRetrievalTimeout 

type ErrRetrievalTimeout struct {
	Duration time.Duration
}

ErrRetrievalTimeout is returned when the retrieval pipeline exceeds its deadline.

func (ErrRetrievalTimeout) Error 

func (e ErrRetrievalTimeout) Error() string

type ErrSymbolNotFound 

type ErrSymbolNotFound struct {
	QualifiedName string
}

ErrSymbolNotFound is returned when a specific symbol cannot be located.

func (ErrSymbolNotFound) Error 

func (e ErrSymbolNotFound) Error() string

type ErrTEIUnavailable 

type ErrTEIUnavailable struct {
	URL string
	Err error
}

ErrTEIUnavailable is returned when the Text Embeddings Inference service cannot be reached.

func (ErrTEIUnavailable) Error 

func (e ErrTEIUnavailable) Error() string

func (ErrTEIUnavailable) Unwrap 

func (e ErrTEIUnavailable) Unwrap() error

type ErrWorkspaceNotFound 

type ErrWorkspaceNotFound struct{}

ErrWorkspaceNotFound is returned when a workspace has no indexed code.

func (ErrWorkspaceNotFound) Error 

func (e ErrWorkspaceNotFound) Error() string

type ExtractResult 

type ExtractResult struct {
	FileNode CodeNode   // kind: "file"
	Symbols  []CodeNode // functions, methods, structs, interfaces
	Edges    []CodeEdge // DEFINES edges (file → symbol)
	Calls    []CallSite // function/method call sites
}

ExtractResult holds everything extracted from a single file.

type Extractor 

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

Extractor runs tree-sitter queries against parsed AST to produce CodeNodes and CodeEdges.

func NewExtractor 

func NewExtractor(lang Language) (*Extractor, error)

NewExtractor creates an extractor for a given language.

func NewExtractorFromBytes 

func NewExtractorFromBytes(query []byte) *Extractor

NewExtractorFromBytes creates an extractor from raw query bytes (useful for tests).

func NewExtractorFromBytesWithLang 

func NewExtractorFromBytesWithLang(query []byte, lang Language) *Extractor

NewExtractorFromBytesWithLang creates an extractor from raw query bytes for a specific language.

func (*Extractor) Extract 

func (e *Extractor) Extract(result *ParseResult) (*ExtractResult, error)

Extract runs the query against the parsed tree and builds nodes + edges.

type FalkorClient 

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

FalkorClient manages a connection to FalkorDB for code intelligence operations.

func NewFalkorClient 

func NewFalkorClient(addr, graphName string) (*FalkorClient, error)

NewFalkorClient creates a new client connected to the given FalkorDB instance with the default vector dimension (768). Use NewFalkorClientDim when the embedder produces a different dimension, or the vector index will reject every embedding write.

func NewFalkorClientDim 

func NewFalkorClientDim(addr, graphName string, dim int) (*FalkorClient, error)

NewFalkorClientDim creates a client whose :Code vector index matches dim. The dimension MUST equal the embedder's Dimension() — TEI models vary (bge-base=768, bge-m3=1024, ...) and a mismatch fails all embedding writes.

func (*FalkorClient) Close 

func (c *FalkorClient) Close() error

Close closes the connection to FalkorDB.

func (*FalkorClient) CreateCodeEdge 

func (c *FalkorClient) CreateCodeEdge(_ context.Context, edge CodeEdge) error

CreateCodeEdge creates a direct edge between two :Code nodes.

func (*FalkorClient) CreateCodeEdgeBatch 

func (c *FalkorClient) CreateCodeEdgeBatch(_ context.Context, edges []CodeEdge) error

CreateCodeEdgeBatch creates multiple edges in batches, grouped by edge type.

func (*FalkorClient) DeleteCodeByFile 

func (c *FalkorClient) DeleteCodeByFile(_ context.Context, filePath string) error

DeleteCodeByFile removes all :Code nodes for a given file path. DETACH DELETE also removes all connected edges automatically.

func (*FalkorClient) DeleteOrphanComponents 

func (c *FalkorClient) DeleteOrphanComponents(_ context.Context) (int64, error)

DeleteOrphanComponents removes :Code component nodes that no longer have any symbols attached via PART_OF. This happens when every file in a component's directory is deleted or becomes unsupported — file deletion cascades to symbols (DETACH DELETE) but never touches the directory-keyed component node itself, leaving a ghost that would otherwise appear in GetComponentTree forever. Returns the number of components removed.

func (*FalkorClient) GetCallGraph 

func (c *FalkorClient) GetCallGraph(_ context.Context, qualifiedName, direction string, depth int, maxResults int) (*CallGraph, error)

GetCallGraph traverses CALLS edges forward or reverse from a function.

func (*FalkorClient) GetComponentTree 

func (c *FalkorClient) GetComponentTree(_ context.Context) ([]ComponentNode, error)

GetComponentTree returns all components with their child symbols and dependencies.

func (*FalkorClient) GetComponentsSymbolSummaries 

func (c *FalkorClient) GetComponentsSymbolSummaries(_ context.Context) (map[string][]string, error)

GetComponentsSymbolSummaries returns, for every component in the graph, its CURRENT full symbol list formatted as one-line descriptions. Key: component file_path (the same key EnrichComponents uses).

This exists because the indexer's in-memory `extracted` slice only covers files changed in the current run — enriching from it alone would describe components using partial/stale symbol lists. The graph is the source of truth for what a component contains right now.

func (*FalkorClient) GetFileHashes 

func (c *FalkorClient) GetFileHashes(_ context.Context) (map[string]string, error)

GetFileHashes returns a map[filePath]fileHash for incremental reindexing.

func (*FalkorClient) Graph 

func (c *FalkorClient) Graph() *falkordb.Graph

Graph returns the underlying FalkorDB graph for raw queries.

func (*FalkorClient) Name 

func (c *FalkorClient) Name() string

Name returns the graph name this client operates on.

func (*FalkorClient) Schema 

func (c *FalkorClient) Schema() error

Schema creates all required indexes for :Code nodes.

func (*FalkorClient) SearchCodeBM25 

func (c *FalkorClient) SearchCodeBM25(_ context.Context, queryText, kind string, scopeComponentNames []string, limit int) ([]CodeNode, error)

SearchCodeBM25 performs full-text search over :Code node names. Unlike the ANN path, the fulltext yield is not candidate-capped, so every text match flows through the scope filter before ordering — scoped BM25 has no starvation problem and needs no over-fetch.

func (*FalkorClient) SearchCodeSemantic 

func (c *FalkorClient) SearchCodeSemantic(_ context.Context, embedding []float64, kind string, scopeComponentNames []string, limit int) ([]CodeNode, error)

SearchCodeSemantic performs vector similarity search over :Code nodes. When scopeComponentNames is non-empty, results are restricted to symbols with PART_OF edges into those components. Scope filtering happens AFTER the ANN query (FalkorDB cannot pre-filter vector search), so scoped calls over-fetch internally — see scopeOverfetchK.

func (*FalkorClient) UpdateComponentMeta 

func (c *FalkorClient) UpdateComponentMeta(_ context.Context, filePath, newName, role, description string) error

UpdateComponentMeta updates a component's name, role, description, and qualified_name after LLM enrichment.

The node is matched on file_path — the stable natural key shared with UpsertCodeNode's MERGE — NEVER on the mutable qualified_name. Renaming component A to a name component B will later also receive makes qname-based matching hijack the wrong node (observed live: both api dirs converged onto one node). file_path cannot collide.

func (*FalkorClient) UpsertCodeNode 

func (c *FalkorClient) UpsertCodeNode(_ context.Context, node CodeNode) (string, error)

UpsertCodeNode creates or updates a :Code node using its natural key. Nodes with an empty Embedding are stored without the code_embedding property (Cypher NULL assignment removes the attribute).

func (*FalkorClient) UpsertCodeNodeBatch 

func (c *FalkorClient) UpsertCodeNodeBatch(_ context.Context, nodes []CodeNode) ([]string, error)

UpsertCodeNodeBatch creates or updates multiple :Code nodes in batches. Returns the UID for each node in the same order as the input slice.

type GoQualifier 

type GoQualifier struct{}

GoQualifier produces Go-style qualified names: "module/pkg.(*Receiver).Name"

func (*GoQualifier) DiscoverRoot 

func (q *GoQualifier) DiscoverRoot(projectPath string) (string, error)

func (*GoQualifier) Qualify 

func (q *GoQualifier) Qualify(pkg, receiver, name, kind string) string

type IndexRequest 

type IndexRequest struct {
	Path      string
	Languages []string // e.g. ["go", "rust"]
	Ignore    []string // e.g. [".git", "vendor"]
}

IndexRequest configures what to index.

type IndexResult 

type IndexResult struct {
	FilesProcessed int
	NodesCreated   int
	EdgesCreated   int
	Errors         []error
	Duration       time.Duration
}

IndexResult reports what was indexed.

type Indexer 

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

Indexer orchestrates code indexing: walk → parse → extract → qualify → embed → store.

func NewIndexer 

func NewIndexer(client *FalkorClient, embedder Embedder, workers int, opts ...IndexerOption) (*Indexer, error)

NewIndexer creates a new indexer. embedder is required for semantic search.

func (*Indexer) Index 

func (idx *Indexer) Index(ctx context.Context, req IndexRequest) (*IndexResult, error)

Index indexes a directory of source code into FalkorDB.

type IndexerOption 

type IndexerOption func(*Indexer)

IndexerOption configures optional Indexer behavior.

func WithLogger 

func WithLogger(l *slog.Logger) IndexerOption

WithLogger sets a logger for diagnostic output. If nil, the indexer runs silently.

func WithProvider 

func WithProvider(p LLMProvider) IndexerOption

WithProvider enables LLM-based component enrichment during indexing.

type JSQualifier 

type JSQualifier struct{}

JSQualifier produces JS-style qualified names: "src/api.js::handleLogin"

func (*JSQualifier) DiscoverRoot 

func (q *JSQualifier) DiscoverRoot(projectPath string) (string, error)

func (*JSQualifier) Qualify 

func (q *JSQualifier) Qualify(pkg, receiver, name, kind string) string

type LLMEvent 

type LLMEvent struct {
	Type  LLMEventType
	Delta string
	Error error
}

LLMEvent is a single streaming event from an LLM provider.

type LLMEventType 

type LLMEventType string

LLMEventType identifies the kind of streaming event. 

const (
	LLMEventTextDelta LLMEventType = "text_delta"
	LLMEventFinish    LLMEventType = "finish"
	LLMEventError     LLMEventType = "error"
)

type LLMMessage 

type LLMMessage struct {
	Role    string // "user", "assistant", "system"
	Content string
}

LLMMessage is a single message in an LLM conversation.

type LLMParams 

type LLMParams struct {
	Messages    []LLMMessage
	Temperature *float64
	Seed        *int
}

LLMParams configures an LLM completion request.

type LLMProvider 

type LLMProvider interface {
	Stream(ctx context.Context, params LLMParams) (iter.Seq[*LLMEvent], error)
}

LLMProvider is the interface for LLM backends used during component enrichment. Implement this with Ollama, OpenAI, Anthropic, or any other provider.

type Language 

type Language string

Language represents a supported programming language. 

const (
	LangGo         Language = "go"
	LangRust       Language = "rust"
	LangJavaScript Language = "javascript"
	LangPython     Language = "python"
)

func DetectLanguage 

func DetectLanguage(ext string) (Language, bool)

DetectLanguage guesses the language from a file extension.

type LanguagePlugin 

type LanguagePlugin struct {
	Name         Language
	Extensions   []string
	NewParser    func() (*Parser, error)
	NewExtractor func() (*Extractor, error)
	Qualifier    SymbolQualifier
}

LanguagePlugin bundles everything needed to support one programming language.

func LookupByExtension 

func LookupByExtension(ext string) (*LanguagePlugin, bool)

LookupByExtension returns the plugin for a file extension (e.g. ".go").

type OllamaProvider 

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

OllamaProvider implements LLMProvider using Ollama's native /api/chat endpoint.

func NewOllamaProvider 

func NewOllamaProvider(baseURL, model, apiKey string) *OllamaProvider

NewOllamaProvider creates an Ollama provider. baseURL: e.g. "http://localhost:11434" or Ollama Cloud URL. model: e.g. "qwen2.5-coder:14b". apiKey: optional Bearer token for authenticated endpoints.

func (*OllamaProvider) Stream 

func (p *OllamaProvider) Stream(ctx context.Context, params LLMParams) (iter.Seq[*LLMEvent], error)

Stream sends a chat request to Ollama and yields text-delta events.

type ParseResult 

type ParseResult struct {
	Tree   *sitter.Tree
	Src    []byte
	Lang   Language
	TsLang *sitter.Language // raw tree-sitter language for query execution
}

ParseResult holds the output of parsing a single file.

type Parser 

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

Parser wraps tree-sitter with language-specific setup.

func NewParser 

func NewParser(lang Language) (*Parser, error)

NewParser creates a parser for the given language.

func (*Parser) Close 

func (p *Parser) Close()

Close releases the underlying tree-sitter parser.

func (*Parser) ParseBytes 

func (p *Parser) ParseBytes(src []byte) (*ParseResult, error)

ParseBytes parses source code from a byte slice (useful for tests).

func (*Parser) ParseFile 

func (p *Parser) ParseFile(filePath string) (*ParseResult, error)

ParseFile reads and parses a single source file.

type PythonQualifier 

type PythonQualifier struct{}

PythonQualifier produces Python-style qualified names: "package.module.ClassName.method"

func (*PythonQualifier) DiscoverRoot 

func (q *PythonQualifier) DiscoverRoot(projectPath string) (string, error)

func (*PythonQualifier) Qualify 

func (q *PythonQualifier) Qualify(pkg, receiver, name, kind string) string

type RerankClient 

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

RerankClient calls a TEI (text-embeddings-inference) cross-encoder /rerank endpoint. A cross-encoder scores each (query, text) pair jointly, giving far sharper relevance than the bi-encoder cosine used for the candidate search — so it pulls the answer-bearing facts to the top, letting us safely keep a focused top-K instead of dumping the full RRF tail.

func NewRerankClient 

func NewRerankClient(url, token string) *RerankClient

NewRerankClient creates a client for a TEI cross-encoder reranking endpoint. The token is optional and used for authenticated endpoints (e.g. Hugging Face Inference).

func (*RerankClient) Rerank 

func (r *RerankClient) Rerank(ctx context.Context, query string, texts []string) ([]int, error)

Rerank returns the input text indices reordered by descending cross-encoder relevance to the query. The returned slice indexes into the original texts.

type RustQualifier 

type RustQualifier struct{}

RustQualifier produces Rust-style qualified names: "crate::module::Name"

func (*RustQualifier) DiscoverRoot 

func (q *RustQualifier) DiscoverRoot(projectPath string) (string, error)

func (*RustQualifier) Qualify 

func (q *RustQualifier) Qualify(pkg, receiver, name, kind string) string

type SymbolQualifier 

type SymbolQualifier interface {
	// Qualify builds a globally unique qualified name.
	// Go:   "github.com/user/repo/internal/api.(*Server).HandleChat"
	// Rust: "crate::auth::service::AuthService::validate"
	// JS:   "src/api.js::handleLogin"
	Qualify(pkg, receiver, name, kind string) string

	// DiscoverRoot finds the project module/package root.
	// Go: reads go.mod → module path
	// Rust: reads Cargo.toml → crate name
	// JS: reads package.json → name (or uses directory)
	DiscoverRoot(projectPath string) (modulePath string, err error)
}

SymbolQualifier builds language-specific qualified names.

type TEIEmbedder 

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

TEIEmbedder calls a Text Embeddings Inference (TEI) service.

func NewTEIEmbedder 

func NewTEIEmbedder(baseURL string, dimension int) (*TEIEmbedder, error)

NewTEIEmbedder creates a TEI embedder with the given endpoint and dimension. dimension must match the model loaded in TEI and the FalkorDB vector index.

func NewTEIEmbedderAuto 

func NewTEIEmbedderAuto(baseURL string) (*TEIEmbedder, error)

NewTEIEmbedderAuto queries TEI for its model dimension and creates an embedder automatically.

func (*TEIEmbedder) Close 

func (e *TEIEmbedder) Close() error

Close closes the HTTP client used by this embedder.

func (*TEIEmbedder) Dimension 

func (e *TEIEmbedder) Dimension() int

Dimension returns the vector dimension this embedder produces.

func (*TEIEmbedder) Embed 

func (e *TEIEmbedder) Embed(ctx context.Context, texts []string) ([][]float32, error)

Embed generates vectors for the given texts. It batches internally.

Source Files

View all Source files

Directories

Path Synopsis
cmd
argus command
argus/ui

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.