mirror of
https://github.com/docker/docker-agent.git
synced 2026-11-04 10:02:15 +00:00
Arnaud Héritier
3c61c72c1c
Extend the detachment pattern already used for the RAG file watcher (context.WithoutCancel) to manager.Initialize, so the 30s toolset-start wait budget (tools.DefaultStartTimeout, shared by all toolset types for #4001 wedged-toolset detection) no longer aborts in-flight indexing. When the caller gives up waiting, indexing keeps running in the background under the single-flight lock and a later turn picks up the toolset once it finishes. Indexing itself is now bounded by a new indexing_timeout config (pkg/config/latest.RAGConfig, default 30m, "0s" = unbounded) instead of the caller's ctx, so a hung provider connection cannot pin the indexing lock forever. Regression tests prove the #4062 backoff-gate invariants hold: a 429 returned by detached indexing still arms the gate, while an indexing_timeout deadline (context.DeadlineExceeded, non-retryable) does not. Part of #4073, depends on #4158 (atomic per-file persistence).
103 lines
3.9 KiB
YAML
103 lines
3.9 KiB
YAML
# This example demonstrates the semantic-embeddings RAG strategy.
|
|
#
|
|
# Unlike chunked-embeddings which embeds raw text chunks directly,
|
|
# semantic-embeddings uses an LLM to generate semantic summaries of each
|
|
# chunk before embedding. This captures the meaning/purpose of code,
|
|
# making retrieval more semantic than direct chunk embedding.
|
|
#
|
|
# Trade-offs:
|
|
# - Higher quality retrieval for code and structured content
|
|
# - Slower indexing (requires LLM call per chunk)
|
|
# - Additional cost from semantic model API calls
|
|
|
|
agents:
|
|
root:
|
|
model: openai/gpt-4o
|
|
description: assistant with semantic code search
|
|
instruction: |
|
|
You are a helpful coding assistant with access to semantic code search.
|
|
Use the search tool to find relevant code based on meaning, not just keywords.
|
|
toolsets:
|
|
- type: rag
|
|
ref: codebase
|
|
|
|
rag:
|
|
codebase:
|
|
tool:
|
|
description: Search the codebase for relevant code snippets by semantic meaning
|
|
# Large knowledge bases can take a while to index. indexing_timeout bounds
|
|
# that single indexing run, detached from the much shorter (30s) toolset-
|
|
# start wait budget: the agent proceeds without this tool on early turns
|
|
# and it appears once indexing finishes. "0s" means no bound; default 30m.
|
|
indexing_timeout: 2h
|
|
docs:
|
|
- ../../pkg/**/*.go
|
|
- ../../cmd/**/*.go
|
|
strategies:
|
|
- type: semantic-embeddings
|
|
# Required: embedding model for vector similarity
|
|
embedding_model: openai/text-embedding-3-small
|
|
vector_dimensions: 1536
|
|
|
|
# Required: chat model to generate semantic summaries of each chunk
|
|
chat_model: openai/gpt-4o-mini
|
|
|
|
# Custom prompt template for generating semantic summaries during indexing.
|
|
# Uses JS template literal syntax with these placeholders:
|
|
# ${path} - full source file path
|
|
# ${basename} - base name of the source file
|
|
# ${chunk_index} - numeric index of the chunk
|
|
# ${content} - raw chunk content
|
|
# ${ast_context} - formatted AST metadata (when ast_context: true)
|
|
semantic_prompt: |
|
|
You are summarizing source code for semantic search.
|
|
|
|
File: ${basename}
|
|
${ast_context}
|
|
|
|
```
|
|
${content}
|
|
```
|
|
|
|
In 2-4 sentences, explain what this code does. Be specific:
|
|
- Name exact functions, types, and methods
|
|
- Mention key dependencies or libraries used
|
|
- Describe inputs, outputs, and notable behavior
|
|
|
|
# Optional: database path (defaults to auto-generated name)
|
|
database: ./semantic_embeddings.db
|
|
|
|
# Optional: similarity settings
|
|
similarity_metric: cosine_similarity
|
|
threshold: 0.3
|
|
limit: 10
|
|
|
|
# Optional: performance tuning
|
|
embedding_batch_size: 50 # chunks per embedding API call
|
|
max_embedding_concurrency: 3 # parallel embedding/LLM requests
|
|
max_indexing_concurrency: 3 # parallel file indexing
|
|
|
|
# Optional: include AST metadata in semantic prompt (best with code_aware chunking)
|
|
ast_context: true
|
|
|
|
# Optional: chunking configuration
|
|
chunking:
|
|
size: 1000
|
|
respect_word_boundaries: true
|
|
code_aware: true # Use tree-sitter for AST-based chunking
|
|
|
|
results:
|
|
# Optional: rerank results using an LLM for better relevance
|
|
reranking:
|
|
model: openai/gpt-4o-mini
|
|
threshold: 0.3
|
|
# Custom criteria to guide the reranking model's relevance scoring
|
|
criteria: |
|
|
When scoring relevance, prioritize:
|
|
- Code that directly implements the queried functionality
|
|
- Functions and methods over comments or documentation
|
|
- Complete implementations over partial snippets
|
|
deduplicate: true
|
|
return_full_content: false # return full document content instead of just the matched chunks
|
|
limit: 5
|