refactor: replace requests with httpx

[isaacbmiller]

Summary

Replace requests with httpx across all 7 source files and 1 test file. httpx is already a transitive dependency via openai and litellm, so this adds no new packages while removing requests, urllib3, and charset-normalizer from the dependency tree.

Changes

File	Change
dspy/dsp/colbertv2.py	requests.get/post -> httpx.get/post
dspy/utils/__init__.py	Streaming download: requests.get(stream=True) + iter_content -> httpx.stream('GET') + iter_bytes, allow_redirects -> follow_redirects
dspy/adapters/types/image.py	requests.get -> httpx.get
dspy/adapters/types/audio.py	requests.get -> httpx.get
dspy/retrievers/databricks_rm.py	requests.post -> httpx.post
dspy/clients/databricks.py	requests.get/put/post -> httpx.get/put/post, removed no-op json body from GET request
dspy/clients/lm_local.py	requests.get -> httpx.get, requests.exceptions.RequestException -> httpx.HTTPError
pyproject.toml	requests>=2.31.0 -> httpx>=0.23.0

Notes

• httpx API is nearly 1:1 with requests for these use cases
• httpx>=0.23.0 matches openai's own lower bound
• Net removal: ~783 KB (requests 199KB + urllib3 409KB + charset-normalizer 175KB)

[greptile-apps]

Greptile Summary

Replaces requests with httpx across 7 source files and 2 test files. Since httpx is already a transitive dependency via openai and litellm, this removes requests, urllib3, and charset-normalizer (~783 KB) with no new packages added.

• All production call sites correctly add follow_redirects=True (addressing the redirect regression noted in earlier review rounds) and the streaming download in utils/__init__.py gains an explicit timeout=300.
• Three httpx.get() calls in tests/signatures/test_adapter_image.py are missing follow_redirects=True, unlike every other call site in the PR; the test_mime_type_from_response_headers test is the most sensitive because it compares an expected MIME type fetched without redirect-following against encode_image() which does follow redirects.
• The GET request in clients/databricks.py silently drops the previously-sent json={\"name\": model_name} body; this is likely harmless since the name is in the URL path, but worth confirming against the Databricks API contract.

Confidence Score: 5/5

Safe to merge; all production HTTP call sites are correctly migrated and the previously flagged redirect and timeout gaps have been addressed.

Every source file change is a clean mechanical substitution with follow_redirects=True and appropriate timeouts. The only gaps are three missing follow_redirects=True flags in test helpers and a clarification question about the removed GET JSON body — neither affects production behaviour.

tests/signatures/test_adapter_image.py (missing follow_redirects in three httpx.get calls) and dspy/clients/databricks.py (removed JSON body from GET request)

Important Files Changed

Filename	Overview
dspy/clients/lm_local.py	Swaps requests.exceptions.RequestException for httpx.HTTPError; correct because ConnectError inherits from RequestError which inherits from HTTPError, so the retry loop still catches connection-refused errors.
dspy/utils/init.py	Streaming download migrated to httpx.stream; explicit timeout=300 prevents mid-stream ReadTimeout on large/slow downloads that the previous thread flagged.
dspy/clients/databricks.py	All three HTTP calls gain follow_redirects=True; the no-op JSON body was removed from the GET request — confirm Databricks API doesn't read it.
tests/signatures/test_adapter_image.py	Three httpx.get() calls are missing follow_redirects=True, unlike all production call sites; could cause test/expected-value mismatch if URLs redirect.
dspy/adapters/types/audio.py	Clean swap to httpx.get with follow_redirects=True added, addressing previously flagged regression.
dspy/adapters/types/image.py	Clean swap to httpx.get with follow_redirects=True and SSL verify parameter preserved.
dspy/dsp/colbertv2.py	Both GET and POST calls migrated cleanly; existing timeout=10 preserved and follow_redirects=True added.
dspy/retrievers/databricks_rm.py	Single POST call migrated with follow_redirects=True; no functional changes to error handling or response parsing.
pyproject.toml	Swaps requests>=2.31.0 for httpx>=0.23.0, matching OpenAI's own lower bound; no new transitive dependencies added.
tests/retrievers/test_colbertv2.py	All four patch targets updated from requests.get/post to httpx.get/post; tests remain correct.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A[HTTP call site] --> B{follow_redirects=True?}
    B -->|YES - all prod files| C[httpx follows 3xx chain]
    B -->|NO - 3 test helpers| D[httpx stops at first 3xx]
    C --> E[raise_for_status and json operate on final response]
    D --> F[raise_for_status passes - 3xx is not an error]
    F --> G[content and headers come from redirect body not target]

    style D fill:#ffe0b2
    style G fill:#ffcccc

Loading

_{Reviews (4): Last reviewed commit: "refactor: replace requests with httpx" | Re-trigger Greptile}