retrievalagent 0.2.9

retrievalagent — multi-backend retrieval-augmented generation with LangGraph

retrievalagent

An autonomous retrieval-augmented generation agent. Plug in any vector store, any LLM, any reranker. Hybrid search, reranking, query rewriting, an LLM quality gate, and an autonomous retry loop — built on LangGraph.

---

from retrievalagent import init_agent

rag = init_agent("documents", model="openai:gpt-5.4", backend="qdrant")
state = rag.chat("What is the status of operation overlord?")
print(state.answer)

---

Scope — Retrieval, Not Ingestion

retrievalagent is built for retrieval quality at query time — hybrid search, reranking, query rewriting, an autonomous retry loop, and an LLM quality gate.

Ingestion is out of scope. The library does not chunk, clean, embed-at-scale, or index your corpus. Use a dedicated tool for that — Docling, Unstructured, LlamaIndex, a Databricks job, or a custom script — then point retrievalagent at the resulting index. Every backend exposes a minimal add_documents() helper for convenience and smoke tests; it is not meant to replace a real ingestion pipeline.

The narrow surface is deliberate: one thing, done well.

---

What it does

Most retrieval systems do a single search pass. retrievalagent runs a state machine that retrieves, evaluates the result, rewrites when needed, and retries — all autonomously, up to max_iter rounds.

Per query the agent will:

1. Understand the intent — rewrite the question into precise search keywords, detect keyword-vs-semantic, and pick the hybrid ratio.
2. Search broadly — run query variants in parallel across BM25 and vector search; fuse the results; rerank.
3. Evaluate — an LLM quality gate decides whether the retrieved docs actually answer the question.
4. Adapt — if not, rewrite the query and retry; on hard failures, swarm-retrieve with parallel strategies as a fallback.
5. Generate — only once the evidence holds, produce a cited, grounded answer.

---

Features

• Fully async pipeline — parallel HyDE + preprocessing, zero blocking calls; every public op has sync and async variants.
• LLM quality gate — rejects weak results, drives the rewrite loop until the evidence holds.
• Multi-query swarm — fans out across BM25 + vector simultaneously, fuses results.
• Autonomous retry loop — retrieve → judge → rewrite → retry, up to max_iter rounds.
• Hybrid search — BM25 + vector, fused with RRF or DBSF.
• HyDE — hypothetical document embeddings for vague queries.
• Tool-calling agent — get_index_settings, get_filter_values, search_hybrid, search_bm25, rerank_results; the LLM picks tools dynamically.
• Multiple rerankers — Cohere, HuggingFace, Jina, ColBERT, RankGPT, embed-anything, or a custom callable.
• 8 search backends — Meilisearch, Azure AI Search, ChromaDB, LanceDB, Qdrant, pgvector, DuckDB, InMemory.
• Any LLM — OpenAI, Azure, Anthropic, Ollama, Vertex AI, or any LangChain BaseChatModel.
• One-line init — init_agent("docs", model="openai:gpt-5.4", backend="qdrant").
• Multi-turn chat — conversation history with citation-aware answers.
• Auto-strategy — the agent samples your collection at init and tunes itself.
• Optuna auto-tuner — 20+ retrieval knobs tuned to your corpus in ~5 min; None is a first-class value for disabling stages. Full guide.

---

Install

# Recommended — Meilisearch + Cohere reranker + interactive CLI
pip install retrievalagent[recommended]

# Base only — in-memory backend, BM25 keyword search
pip install retrievalagent

Extra	What you get	Command
recommended	Meilisearch + Cohere reranker + Rich CLI	pip install retrievalagent[recommended]
cli	Interactive CLI with guided setup wizard	pip install retrievalagent[cli]
all	Every backend + reranker + CLI	pip install retrievalagent[all]

Individual backends & rerankers

pip install retrievalagent[meilisearch]
pip install retrievalagent[azure]
pip install retrievalagent[chromadb]
pip install retrievalagent[lancedb]
pip install retrievalagent[pgvector]
pip install retrievalagent[qdrant]
pip install retrievalagent[duckdb]
pip install retrievalagent[cohere]
pip install retrievalagent[huggingface]
pip install retrievalagent[jina]
pip install retrievalagent[rerankers]      # ColBERT, Flashrank, RankGPT, …
pip install retrievalagent[embed-anything] # Local Rust-accelerated embeddings + reranking

Mix and match: pip install retrievalagent[qdrant,cohere,cli]

---

Quick Start

One-liner with init_agent

The fastest way to get started — no provider imports, string aliases for everything:

from retrievalagent import init_agent

# Minimal — in-memory backend, LLM from env vars
rag = init_agent("docs")

# OpenAI + Qdrant + Cohere reranker
rag = init_agent(
    "my-collection",
    model="openai:gpt-5.4",
    backend="qdrant",
    backend_url="http://localhost:6333",
    reranker="cohere",
)

# Anthropic + Azure AI Search (native vectorisation, no client-side embeddings)
rag = init_agent(
    "my-index",
    model="anthropic:claude-sonnet-4-6",
    gen_model="anthropic:claude-opus-4-6",
    backend="azure",
    backend_url="https://my-search.search.windows.net",
    reranker="huggingface",
    auto_strategy=True,
)

# Fully local — Ollama + ChromaDB + HuggingFace cross-encoder
rag = init_agent(
    "docs",
    model="ollama:llama3",
    backend="chroma",
    reranker="huggingface",
    reranker_model="cross-encoder/ms-marco-MiniLM-L-6-v2",
)

Multi-collection routing

Pass several collections and let the agent decide which to search. The LLM picks the relevant subset before retrieval, using either the collection names alone or optional natural-language descriptions.

from retrievalagent import init_agent

# List form — LLM routes by name only
rag = init_agent(
    collections=["products", "faq", "policies"],
    backend="qdrant",
    backend_url="http://localhost:6333",
    model="openai:gpt-5.4",
)

# Dict form — LLM routes using descriptions (better precision)
rag = init_agent(
    collections={
        "products": "Product catalog: SKUs, prices, specs, availability",
        "faq":      "Customer-facing FAQ, troubleshooting, return policy",
        "policies": "Internal HR/legal/compliance policy documents",
    },
    backend="qdrant",
    backend_url="http://localhost:6333",
    model="openai:gpt-5.4",
)

rag.invoke("What's our return policy?")       # → routes to faq / policies
rag.invoke("Price of SKU 12345?")              # → routes to products

Each retrieved document carries its origin in metadata["_collection"] so you can merge, filter, or attribute citations downstream. One backend instance is built per collection; they share the same backend type and URL.

Backend aliases

Alias	Class	Extra
"memory" / "in_memory"	InMemoryBackend	(none)
"chroma" / "chromadb"	ChromaDBBackend	retrievalagent[chromadb]
"qdrant"	QdrantBackend	retrievalagent[qdrant]
"lancedb" / "lance"	LanceDBBackend	retrievalagent[lancedb]
"duckdb"	DuckDBBackend	retrievalagent[duckdb]
"pgvector" / "pg"	PgvectorBackend	retrievalagent[pgvector]
"meilisearch"	MeilisearchBackend	retrievalagent[meilisearch]
"azure"	AzureAISearchBackend	retrievalagent[azure]

Reranker aliases

Alias	Class	reranker_model	Extra
"cohere"	CohereReranker	Cohere model name (default: rerank-v3.5)	retrievalagent[cohere]
"huggingface" / "hf"	HuggingFaceReranker	HF model name (default: cross-encoder/ms-marco-MiniLM-L-6-v2)	retrievalagent[huggingface]
"jina"	JinaReranker	Jina model name (default: jina-reranker-v2-base-multilingual)	retrievalagent[jina]
"llm"	LLMReranker	(uses the agent's LLM)	(none)
"rerankers"	RerankersReranker	Any model from the rerankers library	retrievalagent[rerankers]
"embed-anything"	EmbedAnythingReranker	ONNX reranker model (default: jina-reranker-v1-turbo-en)	retrievalagent[embed-anything]

# Cohere (default model)
rag = init_agent("docs", model="openai:gpt-5.4", reranker="cohere")

# HuggingFace — multilingual model
rag = init_agent("docs", model="openai:gpt-5.4", reranker="huggingface",
                 reranker_model="cross-encoder/mmarco-mMiniLMv2-L12-H384-v1")

# Jina
rag = init_agent("docs", model="openai:gpt-5.4", reranker="jina")  # uses JINA_API_KEY

# ColBERT via rerankers library
rag = init_agent("docs", model="openai:gpt-5.4", reranker="rerankers",
                 reranker_model="colbert-ir/colbertv2.0",
                 reranker_kwargs={"model_type": "colbert"})

# Pass a pre-built reranker instance directly
from retrievalagent import CohereReranker
rag = init_agent("docs", reranker=CohereReranker(model="rerank-v3.5", api_key="..."))

Model strings: any "provider:model-name" from LangChain's init_chat_model — openai, anthropic, azure_openai, google_vertexai, ollama, groq, mistralai, and more

Manual setup

from retrievalagent import Agent, InMemoryBackend

backend = InMemoryBackend(embed_fn=my_embed_fn)
backend.add_documents([
    {"content": "RAG combines retrieval with generation", "source": "wiki"},
    {"content": "Vector search finds similar embeddings", "source": "docs"},
])

rag = Agent(index="demo", backend=backend)

# Single query → full answer
state = rag.invoke("What is retrieval-augmented generation?")
print(state.answer)

# Retrieve only — documents without LLM answer
query, docs = rag.retrieve_documents("What is retrieval-augmented generation?")
for doc in docs:
    print(doc.page_content)

# Override top-K at call time
query, docs = rag.retrieve_documents("hybrid search", top_k=3)

Agent.from_model — model string with explicit backend

from retrievalagent import Agent, QdrantBackend

rag = Agent.from_model(
    "openai:gpt-5.4-mini",          # fast model for routing & rewriting
    index="docs",
    gen_model="openai:gpt-5.4",     # powerful model for the final answer
    backend=QdrantBackend("docs", url="http://localhost:6333"),
)

---

Multi-turn Chat

from retrievalagent import Agent, ConversationTurn

rag = Agent(index="articles")
history: list[ConversationTurn] = []

state = rag.chat("What is hybrid search?", history)
history.append(ConversationTurn(question="What is hybrid search?", answer=state.answer))

state = rag.chat("How does it compare to pure vector search?", history)
print(state.answer)
print(f"Sources: {len(state.documents)}")

Async variant:

state = await rag.achat("What is hybrid search?", history)

Search-knowledge memory with mem0

history= only carries the current session. For long-term search knowledge that improves retrieval on the same corpus over time, plug mem0 into the agent. The store grows into a corpus-aware glossary of term mappings the agent has learned — informal-to-formal terms, brand spellings, aliases, common typos. It is not a user-preferences store.

When a user query resolves through a non-trivial term expansion (the matching documents used a different surface form than the query), the agent's grader flags it for storage. On future queries, mem0 recalls the relevant mapping and feeds it into BM25 so the same expansion happens automatically.

pip install mem0ai

from retrievalagent import init_agent

rag = init_agent("articles", memory=True)

cfg = {"configurable": {"user_id": "alice"}}
rag.invoke("...", config=cfg)
# → grader may store a search-fact (synonym / alias / typo mapping)
rag.invoke("...", config=cfg)
# → if a stored mapping clears the relevance gate it is injected
#   into BM25 and the system prompt

Two thresholds gate the flow:

• memory_relevance_threshold (env RAG_MEMORY_RELEVANCE_THRESHOLD, default 0.7) — mem0 cosine score the recall must exceed before a stored fact reaches retrieval/generation.
• memory_storage_threshold (env RAG_MEMORY_STORAGE_THRESHOLD, default 0.85) — LLM memory_confidence the grader must report before a new fact is persisted.

Writes are fire-and-forget: the graph schedules mem0.add(...) as a background asyncio task; the user-facing response never waits on memory I/O. await rag.adrain_background() before shutdown if you need the writes to land before exit.

state.trace carries the decision events (read_memory with memories/n_kept/threshold or skipped: below_threshold; final_grade with memory_should_store/memory_confidence) so you can tune the thresholds for your corpus. See docs/memory.md for the full memory matrix (history vs. checkpointer vs. memory_store vs. mem0).

---

Architecture

retrievalagent has two operating modes — both fully autonomous:

Graph mode (rag.chat / rag.invoke)

The default. A LangGraph state machine that runs the full agentic pipeline:

Query
  │
  ├─[HyDE]──────────────────────────────────────────┐
  │  Hypothetical document embedding (parallel)      │
  │                                                  ▼
  ▼                                         [Embed HyDE text]
[Preprocess]                                         │
  Extract keywords + variants                        │
  Detect semantic_ratio + fusion strategy            │
  │                                                  │
  └──────────────────────────────────────────────────┘
                        │
                        ▼
              [Hybrid Search × N queries]
               BM25 + Vector, multi-arm
                        │
                        ▼
               [RRF / DBSF Fusion]
                        │
                        ▼
                    [Rerank]
               Cohere / HF / Jina / embed-anything / LLM
                        │
                        ▼
               [Quality Gate]
               LLM judges relevance
                   │         │
                (good)     (bad)
                   │         │
                   ▼         ▼
              [Generate]  [Rewrite] ──► loop (max_iter)
                   │
                   ▼
        Answer + [n] inline citations

Tool-calling agent mode (rag.invoke_agent)

The agent receives a set of tools and reasons step-by-step, calling them in whatever order makes sense for the question. No fixed pipeline — pure field improvisation:

Query
  │
  ▼
[LLM Agent]  ◄──────────────────────────────────────┐
  Thinks: "What do I need to answer this?"           │
  │                                                  │
  ├── get_index_settings()                           │
  │   Discover filterable / sortable / boost fields  │
  │                                                  │
  ├── get_filter_values(field)                       │
  │   Sample real stored values for a field          │
  │   → build precise filter expressions             │
  │                                                  │
  ├── search_hybrid(query, filter, sort_fields)      │
  │   BM25 + vector, optional filter + sort boost    │
  │                                                  │
  ├── search_bm25(query, filter)                     │
  │   Fallback pure keyword search                   │
  │                                                  │
  ├── rerank_results(query, hits)                    │
  │   Re-rank with configured reranker               │
  │                                                  │
  └── [needs more info?] ─────────────────────────► │

  [done]
  │
  ▼
Answer  (tool calls explained inline)

Use invoke_agent when questions involve dynamic filtering — the agent inspects the index schema, samples real field values, builds filters on the fly, and decides whether to sort by business signals like popularity or recency.

---

Examples

1. Knowledge base Q&A (InMemory, no external services)

from retrievalagent import AgenticRAG, InMemoryBackend
from langchain_openai import ChatOpenAI

docs = [
    {"id": "1", "content": "The Eiffel Tower was built in 1889 for the World's Fair in Paris.", "topic": "history"},
    {"id": "2", "content": "The Louvre is the world's largest art museum, located in Paris.", "topic": "art"},
    {"id": "3", "content": "Photosynthesis converts sunlight and CO2 into glucose and oxygen.", "topic": "science"},
    {"id": "4", "content": "The Python programming language was created by Guido van Rossum in 1991.", "topic": "tech"},
    {"id": "5", "content": "Machine learning is a subset of artificial intelligence.", "topic": "tech"},
]

backend = InMemoryBackend(documents=docs)
llm = ChatOpenAI(model="gpt-5.4-mini")

rag = AgenticRAG(index="kb", backend=backend, llm=llm, gen_llm=llm)

state = rag.invoke("When was the Eiffel Tower built?")
print(state.answer)
# → "The Eiffel Tower was built in 1889 for the World's Fair in Paris. [1]"
print(state.query)        # rewritten query
print(state.iterations)   # how many retrieval rounds it took

---

2. Retrieve documents without generating an answer

Useful when you want the docs and will handle the answer yourself:

from retrievalagent import AgenticRAG, InMemoryBackend

rag = AgenticRAG(index="kb", backend=backend)

query, docs = rag.retrieve_documents("machine learning", top_k=3)
print(f"Rewritten query: {query}")
for doc in docs:
    print(doc.page_content)
    print(doc.metadata)  # original fields + _rankingScore

---

3. Multi-turn chat

from retrievalagent import AgenticRAG, InMemoryBackend, ConversationTurn

rag = AgenticRAG(index="kb", backend=backend, llm=llm, gen_llm=llm)
history: list[ConversationTurn] = []

q1 = "What is machine learning?"
s1 = rag.chat(q1, history)
history.append(ConversationTurn(question=q1, answer=s1.answer))
print(s1.answer)

q2 = "How does it relate to AI?"   # pronoun resolved from history
s2 = rag.chat(q2, history)
history.append(ConversationTurn(question=q2, answer=s2.answer))
print(s2.answer)

---

4. Always-on filter (e-commerce: in-stock items only)

from retrievalagent import AgenticRAG, MeilisearchBackend

backend = MeilisearchBackend(
    "products",
    url="http://localhost:7700",
    api_key="masterKey",
)

# Every search is scoped to in-stock items — no per-call boilerplate
rag = AgenticRAG(
    index="products",
    backend=backend,
    filter="is_in_stock = true",
    llm=llm,
    gen_llm=llm,
)

state = rag.invoke("red running shoes size 42")
for doc in state.documents:
    print(doc.metadata["product_name"], "|", doc.metadata["price"])

---

5. Filter + own-brand exclusion

# Exclude own-brand articles and search for third-party alternatives
rag = AgenticRAG(
    index="products",
    backend=backend,
    filter="is_own_brand = false",
    llm=llm,
    gen_llm=llm,
)

state = rag.invoke("Find alternatives to our house-brand brake cleaner 500ml")
print(state.answer)
# LLM strips the brand prefix, rewrites to "brake cleaner 500ml",
# filter ensures only third-party results are returned.

---

6. Async usage (FastAPI / Databricks / Jupyter)

import asyncio
from retrievalagent import AgenticRAG, InMemoryBackend

rag = AgenticRAG(index="kb", backend=backend, llm=llm, gen_llm=llm)

# Async single query
state = await rag.ainvoke("What is photosynthesis?")
print(state.answer)

# Async batch — runs all queries in parallel
states = await rag.abatch([
    "What is photosynthesis?",
    "Who created Python?",
    "Where is the Louvre?",
])
for s in states:
    print(s.answer)

Sync variants work from any context including Databricks/Jupyter (running event loop is handled automatically):

# Safe to call from a notebook cell even with a running event loop
state = rag.invoke("What is photosynthesis?")
states = rag.batch(["question one", "question two"])

---

7. Tool-calling agent — dynamic filter discovery

When you don't know the filter values upfront, the agent inspects the schema and samples field values itself:

from retrievalagent import AgenticRAG, MeilisearchBackend

rag = AgenticRAG(
    index="products",
    backend=MeilisearchBackend("products", url="http://localhost:7700"),
    llm=llm,
    gen_llm=llm,
)

# Agent calls get_index_settings() → get_filter_values("brand") →
# search_hybrid(filter="brand = 'Bosch'", sort_fields=["popularity"])
result = rag.invoke_agent("Show me the most popular Bosch power tools")
print(result)

---

8. Streaming the final answer

async def stream_answer():
    async for chunk in rag.astream("Explain hybrid search in simple terms"):
        print(chunk, end="", flush=True)

asyncio.run(stream_answer())

---

9. Qdrant — vector search with metadata filter

from retrievalagent import AgenticRAG, QdrantBackend
from qdrant_client import QdrantClient, models

# Insert docs (done once)
client = QdrantClient("http://localhost:6333")
client.upsert("articles", points=[
    models.PointStruct(id=1, vector=embed("RAG combines retrieval and generation"),
                       payload={"content": "RAG combines retrieval and generation", "year": 2023}),
    models.PointStruct(id=2, vector=embed("Vector databases store high-dimensional embeddings"),
                       payload={"content": "Vector databases store high-dimensional embeddings", "year": 2022}),
])

from qdrant_client.models import FieldCondition, MatchValue

rag = AgenticRAG(
    index="articles",
    backend=QdrantBackend("articles", url="http://localhost:6333", embed_fn=embed),
    llm=llm,
    gen_llm=llm,
)

# Pass native Qdrant filter dict — no string translation needed
state = rag.invoke("what is RAG?")
# Or with explicit filter at retrieve time:
_, docs = rag.retrieve_documents("vector databases")

---

10. Custom instructions (tone / domain)

rag = AgenticRAG(
    index="legal_docs",
    backend=backend,
    llm=llm,
    gen_llm=llm,
    instructions=(
        "You are a legal assistant. Answer in formal language. "
        "Always cite the article number when referencing a law. "
        "If the context is insufficient, say so explicitly."
    ),
)

state = rag.invoke("What are the notice periods for dismissal?")
print(state.answer)

---

Backends

Azure AI Search

Native hybrid search — no client-side embeddings needed when the index has an integrated vectorizer:

from retrievalagent import Agent, AzureAISearchBackend

# Native vectorization — service embeds the query server-side
rag = Agent(
    index="my-index",
    backend=AzureAISearchBackend(
        "my-index",
        endpoint="https://my-search.search.windows.net",
        api_key="...",
    ),
)

# Client-side vectorization
rag = Agent(
    index="my-index",
    backend=AzureAISearchBackend(
        "my-index",
        endpoint="https://my-search.search.windows.net",
        api_key="...",
        embed_fn=my_embed_fn,
    ),
)

# With Azure semantic reranking
rag = Agent(
    index="my-index",
    backend=AzureAISearchBackend(
        "my-index",
        endpoint="https://my-search.search.windows.net",
        api_key="...",
        semantic_config="my-semantic-config",
    ),
)

Qdrant

from retrievalagent import Agent, QdrantBackend

rag = Agent(
    index="my_collection",
    backend=QdrantBackend("my_collection", url="http://localhost:6333", embed_fn=my_embed_fn),
)

ChromaDB

from retrievalagent import Agent, ChromaDBBackend

rag = Agent(
    index="my_collection",
    backend=ChromaDBBackend("my_collection", path="./chroma_db", embed_fn=my_embed_fn),
)

LanceDB

from retrievalagent import Agent, LanceDBBackend

rag = Agent(
    index="docs",
    backend=LanceDBBackend("docs", db_uri="./lancedb", embed_fn=my_embed_fn),
)

PostgreSQL + pgvector

from retrievalagent import Agent, PgvectorBackend

rag = Agent(
    index="documents",
    backend=PgvectorBackend(
        "documents",
        dsn="postgresql://user:pass@localhost:5432/mydb",
        embed_fn=my_embed_fn,
    ),
)

DuckDB

from retrievalagent import Agent, DuckDBBackend

rag = Agent(
    index="vectors",
    backend=DuckDBBackend("vectors", db_path="./my.duckdb", embed_fn=my_embed_fn),
)

Meilisearch

from retrievalagent import Agent, MeilisearchBackend

rag = Agent(
    index="articles",
    backend=MeilisearchBackend("articles", url="http://localhost:7700", api_key="masterKey"),
)

InMemory (default, zero dependencies)

from retrievalagent import Agent, InMemoryBackend

backend = InMemoryBackend(embed_fn=my_embed_fn)
backend.add_documents([
    {"content": "RAG combines retrieval with generation", "source": "wiki"},
    {"content": "Vector search finds similar embeddings", "source": "docs"},
])

rag = Agent(index="demo", backend=backend)

---

LLM Configuration

Pass a pre-built LangChain model or use init_agent / Agent.from_model for string-based init.
When using Agent directly, configure via env vars or pass an explicit model instance.

OpenAI

from langchain_openai import ChatOpenAI
from retrievalagent import Agent

rag = Agent(
    index="articles",
    llm=ChatOpenAI(model="gpt-5.4", api_key="sk-..."),
    gen_llm=ChatOpenAI(model="gpt-5.4", api_key="sk-..."),
)

Azure OpenAI (explicit keys)

from langchain_openai import AzureChatOpenAI
from retrievalagent import Agent

llm = AzureChatOpenAI(
    azure_endpoint="https://my-resource.openai.azure.com",
    azure_deployment="gpt-5.4",
    api_key="...",
    api_version="2024-12-01-preview",
)
rag = Agent(index="articles", llm=llm, gen_llm=llm)

Azure OpenAI (env vars)

# Set: AZURE_OPENAI_ENDPOINT, AZURE_OPENAI_API_KEY, AZURE_OPENAI_DEPLOYMENT
from retrievalagent import Agent

rag = Agent(index="articles")  # auto-detected

Azure OpenAI with Managed Identity (no API key)

from azure.identity import DefaultAzureCredential, get_bearer_token_provider
from langchain_openai import AzureChatOpenAI
from retrievalagent import Agent

token_provider = get_bearer_token_provider(
    DefaultAzureCredential(), "https://cognitiveservices.azure.com/.default"
)
llm = AzureChatOpenAI(
    azure_endpoint="https://my-resource.openai.azure.com",
    azure_deployment="gpt-5.4",
    azure_ad_token_provider=token_provider,
    api_version="2024-12-01-preview",
)
rag = Agent(index="articles", llm=llm, gen_llm=llm)

Anthropic Claude

pip install langchain-anthropic

from langchain_anthropic import ChatAnthropic
from retrievalagent import Agent

llm = ChatAnthropic(model="claude-sonnet-4-6", api_key="sk-ant-...")
rag = Agent(index="articles", llm=llm, gen_llm=llm)

Ollama (local, no API key)

pip install langchain-ollama

from langchain_ollama import ChatOllama
from retrievalagent import Agent

rag = Agent(
    index="articles",
    llm=ChatOllama(model="llama3.2", base_url="http://localhost:11434"),
    gen_llm=ChatOllama(model="llama3.2", base_url="http://localhost:11434"),
)

Google Vertex AI

pip install langchain-google-vertexai

from langchain_google_vertexai import ChatVertexAI
from retrievalagent import Agent

llm = ChatVertexAI(model="gemini-2.0-flash", project="my-gcp-project", location="us-central1")
rag = Agent(index="articles", llm=llm, gen_llm=llm)

Separate fast and generation models

Use a cheap/fast model for query rewriting and routing, a powerful model for the final answer:

from langchain_openai import AzureChatOpenAI
from retrievalagent import Agent

fast_llm = AzureChatOpenAI(azure_deployment="gpt-5.4-mini", api_key="...", api_version="2024-12-01-preview")
gen_llm  = AzureChatOpenAI(azure_deployment="gpt-5.4",      api_key="...", api_version="2024-12-01-preview")

rag = Agent(index="articles", llm=fast_llm, gen_llm=gen_llm)

---

Rerankers

Cohere

from retrievalagent import Agent, CohereReranker

rag = Agent(index="articles", reranker=CohereReranker(model="rerank-v3.5", api_key="..."))

HuggingFace cross-encoder (local, no API key)

pip install retrievalagent[huggingface]

from retrievalagent import Agent, HuggingFaceReranker

rag = Agent(index="articles", reranker=HuggingFaceReranker())

# Multilingual
rag = Agent(index="articles", reranker=HuggingFaceReranker(model="cross-encoder/mmarco-mMiniLMv2-L12-H384-v1"))

Jina (multilingual API)

pip install retrievalagent[jina]

from retrievalagent import Agent, JinaReranker

rag = Agent(index="articles", reranker=JinaReranker(api_key="..."))  # or JINA_API_KEY env var

rerankers — ColBERT / Flashrank / RankGPT / any cross-encoder

Unified bridge to the rerankers library by answer.ai:

pip install retrievalagent[rerankers]

from retrievalagent import Agent, RerankersReranker

rag = Agent(index="articles", reranker=RerankersReranker("cross-encoder/ms-marco-MiniLM-L-6-v2", model_type="cross-encoder"))
rag = Agent(index="articles", reranker=RerankersReranker("colbert-ir/colbertv2.0", model_type="colbert"))
rag = Agent(index="articles", reranker=RerankersReranker("flashrank", model_type="flashrank"))
rag = Agent(index="articles", reranker=RerankersReranker("gpt-5.4-mini", model_type="rankgpt", api_key="..."))

Embed-anything — Rust-accelerated local embeddings + reranking

Embeddings and reranking in a single Rust-powered package. Fully local — no API keys, no network calls. Powered by embed-anything.

pip install retrievalagent[embed-anything]

from retrievalagent import Agent, EmbedAnythingEmbedder, EmbedAnythingReranker

# Local embeddings — works as embed_fn (callable)
embedder = EmbedAnythingEmbedder("sentence-transformers/all-MiniLM-L6-v2")

# Local reranker — implements Reranker protocol
reranker = EmbedAnythingReranker("jinaai/jina-reranker-v1-turbo-en")

rag = Agent(
    index="articles",
    backend=QdrantBackend("articles", url="http://localhost:6333", embed_fn=embedder),
    embed_fn=embedder,
    reranker=reranker,
)

Mix and match freely — use embed-anything for one piece and a cloud provider for the other:

from retrievalagent import Agent, EmbedAnythingEmbedder, CohereReranker

# Local embeddings + cloud reranker
rag = Agent(index="docs", embed_fn=EmbedAnythingEmbedder(), reranker=CohereReranker())

# Cloud embeddings + local reranker
from retrievalagent import EmbedAnythingReranker
rag = Agent(index="docs", embed_fn=azure_embed_fn, reranker=EmbedAnythingReranker())

Custom reranker

from retrievalagent import Agent, RerankResult

class MyReranker:
    def rerank(self, query: str, documents: list[str], top_n: int) -> list[RerankResult]:
        return [RerankResult(index=i, relevance_score=1.0 / (i + 1)) for i in range(top_n)]

rag = Agent(index="articles", reranker=MyReranker())

---

Tools

When using invoke_agent, the LLM has access to a set of tools it can call in any order. No fixed pipeline — the agent decides what it needs.

Tool	Description
get_index_settings()	Discover filterable, searchable, sortable, and boost fields from the index schema
get_filter_values(field)	Sample real stored values for a field — used to build precise filter expressions
search_hybrid(query, filter_expr, semantic_ratio, sort_fields)	BM25 + vector hybrid search with optional filter and sort boost
search_bm25(query, filter_expr)	Pure keyword search — fallback when hybrid returns poor results
rerank_results(query, hits)	Re-rank a list of hits with the configured reranker

The agent follows this reasoning pattern:

1. Call get_index_settings() to learn the schema
2. If the question names a specific entity, call get_filter_values(field) to find the exact stored value
3. Call search_hybrid() with a filter and/or sort if relevant, otherwise broad hybrid search
4. Fall back to search_bm25() if results are thin
5. Call rerank_results() to surface the most relevant hits
6. Summarise — explaining which filters and signals influenced the answer

from retrievalagent import Agent

rag = Agent(index="products")

# Agent inspects schema, detects brand field, samples values,
# builds filter, sorts by popularity signal — all autonomously
result = rag.invoke_agent("Show me the most popular Bosch power tools")
print(result)

---

Constructor Reference

Agent(
    index="my_index",           # collection / index name
    backend=...,                # SearchBackend (default: InMemoryBackend)
    llm=...,                    # fast LLM — routing, rewrite, filter
    gen_llm=...,                # generation LLM — final answer
    reranker=...,               # Cohere / HuggingFace / Jina / custom
    top_k=10,                   # final result count            [RAG_TOP_K]
    rerank_top_n=5,             # reranker top-n                [RAG_RERANK_TOP_N]
    retrieval_factor=4,         # over-retrieval multiplier     [RAG_RETRIEVAL_FACTOR]
    max_iter=20,                # max retrieve-rewrite cycles   [RAG_MAX_ITER]
    semantic_ratio=0.5,         # hybrid semantic weight        [RAG_SEMANTIC_RATIO]
    fusion="rrf",               # "rrf" or "dbsf"               [RAG_FUSION]
    instructions="",            # extra system prompt for generation
    embed_fn=None,              # (str) -> list[float]
    boost_fn=None,              # (doc_dict) -> float score boost
    filter=None,                # always-on Meilisearch filter expr (e.g. "brand = 'Bosch'")
    category_fields=None,       # fields used by alternative retrieve (None → auto-detect via regex)
    hyde_min_words=8,           # min words to trigger HyDE     [RAG_HYDE_MIN_WORDS]
    hyde_style_hint="",         # style hint for HyDE prompt
    auto_strategy=True,         # auto-tune from index samples
)

Always-on filter

Pin every search to a subset of the index with filter — Meilisearch syntax, AND-joined with any per-call filter (intent, language, ...):

rag = init_agent("products", filter="brand = 'Bosch'")
# every BM25 + vector + hybrid search scoped to Bosch only

The legacy base_filter kwarg still works but emits a DeprecationWarning — migrate to filter at your convenience.

Category fields (alternative retrieve)

The alternative-retrieve fallback broadens the search by pivoting on category-like fields (product groups, taxonomy levels, sections, ...). By default, retrievalagent auto-detects them from the index schema via regex — matching names like category, product_group_l3, article_group_name, kategorie, family, section, ... — and prioritises deeper taxonomy levels (_l3 > _l2 > _l1).

Override explicitly when your schema uses unusual names:

rag = init_agent(
    "products",
    category_fields=["taxonomy_leaf", "taxonomy_parent", "department"],
)

Pass category_fields=[] to disable the fallback entirely.

---

API Reference

Method	Returns	Description
rag.invoke(query)	RAGState	Full RAG pipeline (sync)
rag.ainvoke(query)	RAGState	Full RAG pipeline (async)
rag.chat(query, history)	RAGState	Multi-turn chat (sync)
rag.achat(query, history)	RAGState	Multi-turn chat (async)
rag.retrieve_documents(query, top_k)	(str, list[Document])	Retrieve only, no answer
rag.query(query)	str	Answer string directly
rag.invoke_agent(query)	str	Tool-calling agent mode (sync)
rag.ainvoke_agent(query)	str	Tool-calling agent mode (async)

RAGState fields: answer · documents · query · question · history · iterations

---

Environment Variables

Variable	Description	Default
AZURE_OPENAI_ENDPOINT	Azure OpenAI endpoint	—
AZURE_OPENAI_API_KEY	Azure OpenAI API key	—
AZURE_OPENAI_DEPLOYMENT	Default deployment	—
AZURE_OPENAI_FAST_DEPLOYMENT	Fast model deployment	→ DEPLOYMENT
AZURE_OPENAI_GENERATION_DEPLOYMENT	Generation deployment	→ DEPLOYMENT
AZURE_OPENAI_API_VERSION	API version	2024-12-01-preview
OPENAI_API_KEY	OpenAI API key (fallback)	—
OPENAI_MODEL	OpenAI model name	gpt-5.4
AZURE_COHERE_ENDPOINT	Azure Cohere endpoint	—
AZURE_COHERE_API_KEY	Azure Cohere API key	—
COHERE_API_KEY	Cohere API key (fallback)	—
JINA_API_KEY	Jina reranker API key	—
MEILI_URL	Meilisearch URL	http://localhost:7700
MEILI_KEY	Meilisearch API key	masterKey
RAG_TOP_K	Final result count	10
RAG_RERANK_TOP_N	Reranker top-n	5
RAG_RETRIEVAL_FACTOR	Over-retrieval multiplier	4
RAG_SEMANTIC_RATIO	Hybrid semantic weight	0.5
RAG_FUSION	Fusion strategy	rrf
RAG_HYDE_MIN_WORDS	Min words to trigger HyDE	8

---

Tune It For Your Data

retrievalagent ships with curated tuned defaults in [tool.retrievalagent] of pyproject.toml, found by running the built-in tuner against real German/Swiss product catalog data. These are better than hand-picked defaults for most retrieval tasks.

For peak performance on your corpus (product vs. legal vs. support vs. scientific), run the tuner yourself — it searches 20+ knobs with Optuna TPE sampler and usually beats defaults by 3–10% combined score (hit@5 + paraphrase consistency).

Real benchmark (Meilisearch cloud, 3 German product-catalog indexes, 39 hit cases + 8 paraphrase groups):

Config	hit@5	consistency	stable_top1	combined
Library defaults	0.903	0.750	0.250	0.727
Shipped tuned ([tool.retrievalagent])	0.968	0.792	0.250	0.761
Full corpus-tuned (local)	0.968	0.792	0.375	0.761

combined = hit@5×0.4 + consistency×0.35 + stable_top1×0.25

📘 Full walkthrough: docs/auto-tuning.md — testset design, CLI and Python API, every searched parameter, result interpretation, troubleshooting.

1. Install

pip install 'retrievalagent[tune]'

2. Write a testset

A list of (query, expected_doc_ids, id_field) tuples — or a JSON file:

[
  {"query": "Makita Akku Bohrhammer 18V", "expected_ids": ["SKU-1065144"], "id_field": "sku"},
  {"query": "Bosch Winkelschleifer 125mm", "expected_ids": ["SKU-1057802"], "id_field": "sku"}
]

3. Run the tuner

CLI (fastest path):

python -m retrievalagent.tuner \
  --index my_index \
  --hits testset.json \
  --trials 50 \
  --patience 8         # early-stop if 8 trials show no improvement

Use --pyproject to write to pyproject.toml [tool.retrievalagent] instead of the gitignored retrievalagent.config.toml.

Python API (full control):

from retrievalagent import MeilisearchBackend, RAGConfig
from retrievalagent.tuner import RAGTuner, load_testset
from retrievalagent.utils import _make_azure_embed_fn

tuner = RAGTuner(
    backend_factory=lambda: MeilisearchBackend(index="my_index"),
    embed_fn=_make_azure_embed_fn(),
    hit_cases=load_testset("testset.json"),
    eval_k=5,
    # Optional: let the tuner mix/match weak + thinking models across
    # cost tiers (gen_llm / strong_model stays fixed — quality-critical).
    candidate_models=["azure:gpt-5.4-mini", "azure:gpt-5.4-nano"],
)

best = tuner.tune(
    n_trials=50,
    patience=8,             # early-stop on plateau
    trial_timeout_s=120,    # hung trials score 0, never block the study
)
best.save_toml("retrievalagent.config.toml")   # gitignored — local override (recommended)
# or: best.save_pyproject()          # [tool.retrievalagent] — commit if your team shares tuning

4. Use the tuned config

No code change required — AgenticRAG picks up [tool.retrievalagent] automatically:

from retrievalagent import AgenticRAG, RAGConfig, MeilisearchBackend

rag = AgenticRAG(
    index="my_index",
    backend=MeilisearchBackend("my_index"),
    embed_fn=embed_fn,
    config=RAGConfig.auto(),   # discovers pyproject.toml → retrievalagent.config.toml → env
)

Config discovery order

1. Runtime kwarg — AgenticRAG(config=RAGConfig(...))
2. retrievalagent.config.toml — per-deployment local override. Gitignored by default so your corpus-specific tuning doesn't leak into source control. Wins over pyproject defaults — drop a tuned file here and you're done.
3. [tool.retrievalagent] in pyproject.toml — shipped / shared defaults. retrievalagent ships with curated values here (tuned on real product-catalog data). Matches ruff/black/mypy convention for committed tool config.
4. RAG_* env vars — containers/CI overrides.
5. Library defaults — fallback if nothing else is set.

What gets tuned

Scalar thresholds:

Parameter	Range	Effect
retrieval_factor	2–8	Over-retrieve multiplier before reranking
rerank_top_n	3–10	Docs kept post-rerank
rerank_cap_multiplier	1.5–4	Caps reranker input at top_k × m
semantic_ratio	0.3–0.9	BM25 ⇄ vector balance
fusion	rrf / dbsf	Score fusion strategy
short_query_threshold	3–8	When to skip LLM preprocessing
short_query_sort_tokens	bool	Sort tokens for paraphrase invariance
bm25_fallback_semantic_ratio	0.7–1.0	Semantic ratio used when BM25 fails
rerank_skip_gap	0.05–0.3	Top-1 vs top-5 score gap to skip reranker
name_field_boost_max	0.0–0.5	Post-rerank boost for docs matching query tokens in name_field. Higher → precise lookups win; lower → paraphrase stability.

Optional (None = disable stage, first-class tuning option):

Parameter	Range / None	Effect
bm25_fallback_threshold	0.2–0.6 / None	BM25 score below which we boost semantic. None = never boost.
fast_accept_score	0.5–0.95 / None	BM25 score to accept fast path. None = always slow path.
fast_accept_confidence	0.6–0.95 / None	LLM confidence to accept fast path. None = no LLM confirm.
rerank_skip_dominance	0.6–0.95 / None	Score to skip reranker on obvious hits. None = always rerank.
expert_threshold	0.05–0.3 / None	Gap to escalate to expert reranker. None = never escalate.
hyde_min_words	2–20 / None	Min words to trigger HyDE. None = disable HyDE.

Pipeline-stage toggles:

Parameter	Default	Effect
enable_hyde	true	Hypothetical-document expansion
enable_filter_intent	true	LLM detects filter intent from query
enable_quality_gate	true	LLM judges retrieval quality before answering
enable_preprocess_llm	true	LLM query rewrite + variant generation
enable_reasoning	true	Per-document relevance reasoning (uses thinking_model)

LLM tiers (cost ⇄ quality, mix and match):

Parameter	Role	Typical cheap pick
strong_model	Final cited answer — gen_llm	azure:gpt-5.4
weak_model	Preprocess / quality / filter-intent / rewrite	azure:gpt-5.4-mini
thinking_model	Per-document reasoning when enable_reasoning=true	azure:gpt-5.4-mini

Specs are provider:model — resolved via LangChain's init_chat_model, so every supported provider works (azure:, openai:, anthropic:, bedrock:, ollama:, etc.). None = inherit from env vars.

Optuna's TPE sampler learns from prior trials — 50 trials usually beats hand-tuning. The tuner explores disabling each optional stage as a first-class hypothesis, so it can discover e.g. "your corpus doesn't need filter-intent detection." Use a cheap LLM (gpt-4o-mini) during tuning to keep cost down; swap to production LLM afterwards.

Disabling stages in TOML

TOML has no null, so disabled fields go in a disable list:

[tool.retrievalagent]
top_k = 10
semantic_ratio = 0.5
# disable these optional stages for this corpus:
disable = ["bm25_fallback_threshold", "expert_threshold"]

RAGConfig.from_toml() / from_pyproject() translates disable = [...] into None values on load.

---

CLI

pip install retrievalagent[recommended]

# Guided setup wizard — choose LLM, embedder, backend, reranker
retrievalagent

# Chat mode — full agentic pipeline
retrievalagent --chat -c my_index

# Retriever mode — documents only, no LLM
retrievalagent --retriever -c my_index

# Skip wizard, use env vars
retrievalagent --skip-wizard -c my_index

The wizard guides you through:

1. LLM provider — OpenAI, Anthropic, Ollama, or env default
2. Embedding model — OpenAI, Azure OpenAI, Ollama, or none (BM25 only)
3. Vector store — InMemory, Meilisearch, ChromaDB, Qdrant, pgvector, DuckDB, LanceDB, Azure AI Search
4. Reranker — Cohere, Jina, HuggingFace, LLM-based, or none
5. Mode — Chat (with answers) or Retriever (documents only)

---

License

MIT — Licence to code.