By: the Foundry IQ microsoft-retriever-1 team
Today, we’re introducing Microsoft Retriever 1 an agentic search model trained by Microsoft and available in Foundry IQ. It’s available in preview today, with no provisioning or reindexing.
Key Points
- microsoft-retriever-1 is an agentic search model trained by Microsoft and hosted entirely within Azure's security envelope.
- Agents that delegate retrieval to microsoft-retriever-1 use a third fewer input tokens, with the highest cited recall and answer correctness.
- microsoft-retriever-1 requires no provisioning or re-indexing. Enable it on Foundry IQ search queries or on a knowledge base.
- Adding microsoft-retriever-1 to auto retrieval reasoning effort in knowledge bases cuts typical latency by more than 6x.
- Adding microsoft-retriever-1 to Foundry IQ search queries improves evidence recall by 8-11 points.
Fully managed agentic retrieval in Foundry IQ
In most agentic RAG systems, one large language model manages interactions with the user and drives the retrieval system. It infers information needs, writes search queries, and decides to search again or produce an answer. Every turn of that loop adds a network round trip and fills the context window. Changing models requires tuning the prompts driving the entire system.
With the introduction of Microsoft Retriever 1, Foundry IQ allows you to offload agentic retrieval tasks to an Azure-hosted model. Microsoft Retriever 1 interprets the underlying information need from user queries, reformulates those queries using terminology from documents in your corpus, and returns more relevant results. Enabling Microsoft Retriever 1 in your existing Foundry IQ solution requires no additional provisioning or document ingestion.
In Foundry IQ knowledge bases, auto retrieval reasoning effort replaces the agentic retrieval loop with Microsoft Retriever 1, escalating to a larger language model only when required. After an initial search, Microsoft Retriever 1 reviews the results and decides whether to search again. In our evaluations, only the hardest 20% of queries escalated to that larger model. Compared with a knowledge base using medium retrieval reasoning effort with gpt-6.1-sol, pairing microsoft-retriever-1 with gpt-6.1-sol cut typical latency nearly 6x and gpt-6.1-sol token usage by more than 75%, while retaining 90% of its evidence recall.
How we evaluated Microsoft Retriever 1
We evaluated microsoft-retriever-1 on Foundry IQ using a combination of enterprise and public datasets. For each configuration, we measured retrieval quality and latency then plotted the results. Up and to the left is better: higher quality at lower latency. The line connecting the fastest and most accurate configurations is called the Pareto frontier. microsoft-retriever-1 pushes that frontier outwards and achieves the highest retrieval quality available in Foundry IQ.
In this chart, each point is a knowledge base configuration: a retrieval reasoning effort alone or combined with microsoft-retriever-1. Points higher up have more evidence recall, and points further left have lower median latency. The green line traces the best configurations with microsoft-retriever-1, and the dashed gray line traces the best without it.
The three experiments below cover different ways to use microsoft-retriever-1 with Foundry IQ: knowledge bases, the search API, and larger models delegating agentic retrieval.
Experiment 1: More than 6x lower typical latency for knowledge base retrieval with Microsoft Retriever 1
Foundry IQ knowledge bases offer four retrieval reasoning efforts: minimal, low, medium, and auto. In this release, microsoft-retriever-1 can be run on its own in minimal or be combined with an external LLM in auto. In our evaluations, we used gpt-6.1-sol as the external LLM. Combining auto retrieval reasoning effort with microsoft-retriever-1 retains 90% of the evidence recall of medium while cutting median latency 85% from 5.09 to 0.76 seconds and using 83.6% fewer gpt-6.1-sol tokens. microsoft-retriever-1 was able to handle 80% of the total queries on its own without escalating. In minimal, it raised recall from 56% to 61% with no call to an external LLM.
|
Knowledge base configuration |
GPT-6.1 Sol tokens |
Median latency |
P95 latency |
Evidence recall |
|---|---|---|---|---|
| Medium | 8,873 | 5.09 s | 10.53 s | 71% |
| Auto | 4,831 | 4.46 s | 8.19 s | 69% |
| Auto + microsoft-retriever-1 | 1,453 | 0.76 s | 7.14 s | 64% |
| Low | 836 | 2.48 s | 5.63 s | 61% |
| Minimal + microsoft-retriever-1 | 0 | 0.83 s | 1.56 s | 61% |
| Minimal | 0 | 0.50 s | 0.62 s | 56% |
Experiment 2: Microsoft Retriever 1 raises search API evidence recall by 8 to 16 points
Experiment 1 showed microsoft-retriever-1 inside a knowledge base, where Foundry IQ manages agentic retrieval for you. Many applications query search indexes on Foundry IQ directly, choosing among keyword (BM25) retrieval, vector (embedding) retrieval, or a hybrid of the two, with optional semantic ranking. For these applications, we're introducing a new agentic query type where microsoft-retriever-1 runs an agentic retrieval loop inside a normal search call. In these searches, microsoft-retriever-1 decides which queries to issue and reasons over the results.
We compared search configurations with and without microsoft-retriever-1. Across all evaluations, microsoft-retriever-1 found enough evidence to answer 10.3% of questions that no configuration without microsoft-retriever-1 could.
|
Baseline configuration |
Microsoft-retriever-1-only solves |
Baseline-only solves |
Net gain (p.p.) |
|
BM25 |
17.59% |
1.93% |
+15.66 |
|
BM25 + semantic ranking |
12.33% |
1.87% |
+10.45 |
|
Hybrid (BM25 + vector) |
14.98% |
2.66% |
+12.33 |
|
Hybrid + semantic ranking |
10.65% |
2.52% |
+8.12 |
|
Overall |
10.31% |
2.73% (Any) |
+7.58 |
A question is solved when the retrieved evidence supports a complete grounded answer. Microsoft-retriever-1-only solves are questions solved with microsoft-retriever-1 but not the baseline; baseline-only solves are the reverse. In the Overall row, microsoft-retriever-1-only solves are questions no other configuration solves, and baseline-only solves are questions at least one baseline solves but microsoft-retriever-1 does not.
In search, every microsoft-retriever-1 configuration adds latency, but moves recall well past what the existing configurations reach. Microsoft-retriever-1 with BM25 reaches 64% recall at 1 second, ahead of every configuration without microsoft-retriever-1. Microsoft-retriever-1 with hybrid reaches 67% recall at 1.65 seconds, the highest in the chart.
| Search configuration | Baseline | microsoft-retriever-1 | Delta (p.p.) | Win / tie / loss | Added latency |
|
BM25 |
47.40% |
63.70% |
+16.30 |
28% / 67% / 5% |
+0.72 s |
|
BM25 + semantic ranking |
52.81% |
+10.89 |
19% / 76% / 5% |
+0.61 s | |
|
Hybrid (BM25 + vector) |
54.91% |
67.29% |
+12.39 |
24% / 69% / 7% |
+0.94 s |
|
Hybrid + semantic ranking |
59.32% |
+7.98 |
17% / 76% / 7% |
+0.85 s |
Microsoft Retriever 1 enables semantic ranking by default. Its attributable gain is 8 to 11 p.p. Delta is the change in evidence recall in percentage points (p.p.). Win / tie / loss is the share of question pairs where Microsoft Retriever 1 improved, matched, or reduced recall.
Experiment 3: Agents that delegate retrieval to Microsoft Retriever 1 use 32% fewer input tokens and get more answers right
Every agent needs relevant context. To understand how agents might use agentic retrieval to find that context, we provided gpt-6.1-sol with access to search tools powered by BM25, BM25 with semantic ranking, and BM25 with microsoft-retriever-1. We first required gpt-6.1-sol to use each tool, then let it choose among all three on its own. The system prompt described the technology behind each tool. We then ran our full private and public datasets through each configuration.
| Method | Cited recall | Answer correct | gpt-6.1-sol input tokens | Mean turns | Mean citations |
| BM25 | 68.5% | 63.3% | 9,595 | 1.6 | 5.3 |
| BM25 + semantic ranking | 72.9% | 66.6% | 7,960 (-17%) | 1.5 | 4.6 |
| BM25 + microsoft-retriever-1 | 74.8% | 69.1% | 6,465 (-32%) | 1.4 | 6.0 |
| gpt-6.1-sol choice | 73.0% | 67.6% | 7,960 (-3%) | 1.5 | 5.6 |
When should you use Microsoft Retriever 1?
Our experiments generated over 120,000 agentic retrieval runs. To understand when microsoft-retriever-1 improves recall enough to justify its latency, we used each run to classify its question into one of five query types and measured the difference in evidence recall.
- Factual lookup: Retrieve one directly stated fact. What is the Amber Plan in-network deductible?
- Comparative: Contrast entities or documents. How does Amber's overseas emergency coverage differ from Birch's?
- Multi-hop: Follow a connection across sources. What is the inspection interval for scaffolding holding the Lark water pump?
- Identification: Find a fact or entity based on constraints. Which plan waives specialist referrals and covers remote consultations while traveling abroad?
- Thematic: Surface patterns across the corpus. Looking at the project retrospectives, what problems repeatedly cause schedule delays?
microsoft-retriever-1 gains are largest for identification, multi-hop, and comparative queries:
|
Query type |
BM25 |
BM25 + semantic ranking |
Hybrid |
Hybrid + semantic ranking |
|
Factual lookup |
+10 p.p. 80% → 90% |
+4 p.p. 86% → 90% |
+4 p.p. 90% → 94% |
+1 p.p. 93% → 94% |
|
Comparative |
+12 p.p. 43% → 55% |
+5 p.p. 50% → 55% |
+9 p.p. 52% → 61% |
+3 p.p. 58% → 61% |
|
Multihop |
+12 p.p. 42% → 54% |
+6 p.p. 48% → 54% |
+10 p.p. 49% → 59% |
+4 p.p. 55% → 59% |
|
Identification |
+14 p.p. 39% → 53% |
+8 p.p. 45% → 53% |
+17 p.p. 38% → 55% |
+11 p.p. 44% → 55% |
|
Thematic |
+7 p.p. 20% → 27% |
+4 p.p. 23% → 27% |
+5 p.p. 24% → 29% |
+3 p.p. 26% → 29% |
Each cell shows the change in evidence recall (percentage points) with baseline and microsoft-retriever-1 recall below. Queries are weighted equally within a category. Experiments 1 to 3 average per dataset (macro-averaging) to avoid dataset-size bias, so totals differ from those above.
There are a few key takeaways from this table:
- microsoft-retriever-1 is most valuable for queries that express constraints or require multiple documents.
- Factual lookup queries are already well served by existing search methods.
- Thematic queries remain hard. Many questions require broad traversal of the corpus, and recall remains below 30% in every configuration.
We recommend trying microsoft-retriever-1 in these scenarios:
- Knowledge bases: Enable auto retrieval reasoning effort with microsoft-retriever-1 and compare it with your current retrieval performance. We expect lower latency and improved retrieval quality. If you aren't using an external LLM, use minimal retrieval reasoning effort with microsoft-retriever-1 instead.
- Search indexes: If your queries require reasoning across documents or applying constraints, enable microsoft-retriever-1 on those queries.
- Agents: delegate retrieval to knowledge bases or search indexes with microsoft-retriever-1. We expect reduced token usage and higher evidence call.
Get started with Microsoft Retriever 1
microsoft-retriever-1 is available today in the 2026-11-01-preview API version of Foundry IQ. Enabling it requires setting query type to agentic on a search API request or adding it to the knowledge base model list. Changes are highlighted below:
Using Microsoft Retriever 1 in Foundry IQ search indexes
Set query_type to agentic. No reindexing or additional provisioning is required
results = search.search(
search_text=query,
query_type=QueryType.AGENTIC, # was QueryType.SIMPLE
semantic_configuration_name=semantic_config,
top=20,
)
Using Microsoft Retriever 1 in Foundry IQ knowledge bases
Add microsoft-retriever-1 to the model list in a knowledge base and enable auto or minimal retrieval reasoning effort. No reindexing or additional provisioning is required.
from azure.search.documents.indexes.models import KnowledgeBaseMicrosoftRetriever1Model
api_version = "2026-11-01-preview"
with SearchIndexClient(
endpoint=endpoint,
credential=credential,
api_version=api_version,
) as admin:
baseline = admin.get_knowledge_base("my-existing-kb")
admin.create_or_update_knowledge_base(KnowledgeBase(
name="my-kb-with-msr1",
knowledge_sources=baseline.knowledge_sources,
models=[
KnowledgeBaseMicrosoftRetriever1Model(),
*baseline.models,
],
retrieval_reasoning_effort=KnowledgeRetrievalAutoReasoningEffort(),
output_mode="extractiveData",
))
Learn more by reviewing the documentation.
Appendix
Side by side: Microsoft Retriever 1 vs. general-purpose models.
General purpose LLMs are trained on many tasks: general chat, coding, tool use, and searching the web. microsoft-retriever-1's training objective is to maximize the accuracy and efficiency of agentic retrieval. That singular focus is especially clear when responsive retrieval is critical.
microsoft-retriever-1 recalling evidence for 5 questions vs. a general purpose model.
Benchmark datasets
We evaluated this release on private datasets spanning multiple industries and languages, as well as public datasets
Private datasets include:
- Enterprise collections: Corporate document collections in English, covering domains such as oil and gas, corporate reports, and health insurance member documents.
- MIML Western: A multi-industry multi-lingual corporate document benchmark in English, French, Spanish, and German. We evaluate against the full cross-language, cross-industry index with no a priori filtering.
Public datasets include:
- FDA: English-language FDA drug labels and clinical documents, evaluated as a single search index created from 400-token chunked PDFs
- Natural Questions (Kwiatkowski et al., 2019): A public English open-domain question-answering benchmark built from real, anonymized queries issued to Google Search, with answers annotated from Wikipedia pages.
- HotpotQA (Yang et al., 2018): A public English multi-hop question-answering benchmark over Wikipedia in which answering each question requires finding and reasoning over multiple supporting documents, including comparison questions.
Metric definitions
|
Metric |
What it measures |
|
Evidence recall |
The share of ground-truth nuggets (facts required by the answer) in the documents returned by the relevant API. This is an adaptation of the nugget-based evaluation methodology described in the TREC 2025 RAG Track (Upadhyay et al., 2025). |
|
Cited recall |
Evidence recall over only the documents cited in the answer. It measures how fully the citations cover the required evidence, unlike citation support checks, which only confirm that each citation backs its claim. |
|
Answer Correct |
The share of questions answered fully correctly produced using retrieved evidence. |
|
Win / tie / loss |
The share of question pairs where microsoft-retriever-1 improved, matched, or reduced evidence recall relative to the baseline. |
|
Median latency |
Typical search and retrieve API response time, excluding answer synthesis and grading. |
|
P95 latency |
The search and retrieve API response time that 95% of requests complete within, excluding answer synthesis and grading. This captures tail latency, including queries that auto mode escalates to a larger model. |
|
Added latency |
The extra median latency compared with the baseline configuration. |
|
Input tokens |
The average number of tokens per question, counted with the o200k_base tokenizer. |
Ground-truth nuggets are atomic facts that together fully answer a question.