Haystack installed successfully in the recorded sandbox, but its test command failed. The available tail identifies missing optional libraries in PDF and PowerPoint converter checks. For a Python team evaluating document retrieval, the next useful check is a small pipeline using the files and integrations the application actually needs.
Where explicit pipelines help
Haystack documents modular pipelines and agent workflows with explicit control over retrieval, routing, memory and generation. Its README also describes built-in and custom components, loops, branches and conditional logic.
Consider it when engineers need to inspect several connected steps, such as document ingestion, retrieval and generation. For an application consisting of a single model request, first identify what a pipeline framework would help you manage.
The README lists hosted providers and local models among its integrations. This review does not establish compatibility for a particular combination of models, converters and document stores. Check the dependencies, credentials and services required by the components you select.
What happened when we ran it
The October 6, 2026 sandbox run at commit ff13437 installed dependencies in 19 seconds and built in 2 seconds. The record lists 71 dependencies and 138 MB on disk. These are observations from that environment, not expected setup times or resource requirements for every deployment.
The test command exited 1 after 22 seconds. The parser reported 217 passed, 110 failed, 2 skipped and 90 collection/setup errors. These are unreconciled parser-reported counts: the recorded total of 417 conflicts with the category counts. Complete unique test coverage was not verified, so these figures do not support a pass percentage or a claim about exhaustive upstream coverage.
The failing tail names a PowerPoint converter check and several PDF converter checks. Their ImportErrors identify unavailable optional dependencies pptx and pypdf. That supports a diagnosis for those named checks, not an explanation of the entire failed run. The supplied evidence does not establish that installing those libraries would resolve every failure and error.
Pytest's final line reports 16.68 seconds, distinct from the recorded 22-second outer step. Neither duration measures document-ingestion performance.
The audit record contains zero advisories and zero packages; scan coverage is unknown, so it does not establish that installed dependencies were checked for known vulnerabilities.
The recorded configuration had 3 CPUs, 8 GB RAM and image label ghcr.io/astral-sh/uv:python3.12-bookworm. The image label does not establish actual runtime or operating-system versions. The checkout scan recorded 12,099 files, approximately 151,254 source lines, 106.1 MB and 36 CI workflow files. It found a tests directory and detected no Dockerfile. Those observations do not measure retrieval quality or deployment reliability.
No new execution was performed for this correction. The measured checkout is identified by commit ff13437; this review does not assign it a package release version.
Evaluate the document path you need
If PDF or PowerPoint ingestion is required, use representative files and inspect the extracted text before expanding the application. Provision the selected converter's dependencies in the intended environment, then check retrieval and model output together. This is a proposed acceptance check, not a pipeline executed for this review.
The README points newcomers to an overview, a quick-start guide, tutorials and a cookbook. Those are routes into component selection; the linked walkthroughs were not validated here.
Review serving and telemetry separately
The README points to Hayhooks for wrapping pipelines and agents with custom logic and exposing them through HTTP or MCP. The recorded run did not verify that serving path.
Haystack says it collects anonymous usage statistics when pipeline components are initialized and links to opt-out documentation. Review those settings if your deployment restricts telemetry. The run did not inspect network traffic or verify the contents of telemetry events.
Choosing an alternative
LangGraph describes low-level orchestration for long-running, stateful agents. Its README documents durable execution and resuming after failures. Evaluate it when persistence and resumable execution are central requirements. No comparative reliability or performance was measured here.
LlamaIndex documents an open-source framework for agentic applications, with a core package and selected integrations. Its README says the company's primary focus has shifted toward LlamaParse, liteparse and benchmarking. Evaluate whether the framework's integrations fit your document workflow while distinguishing it from the company's parsing products.
For Haystack, the adoption decision remains open until the document path your application needs works in its intended environment. The named converter errors provide a concrete starting point for that evaluation; the remaining failures and errors are unresolved.