Open
Conversation
- Switch agent/ + mcp/ Dockerfiles from pip to uv (4-6x faster cold builds: agent 30s vs ~150s, mcp 13s vs ~60s; warm rebuilds ~1s). - Rewrite postprovision.sh to be tolerant of local environment quirks: use a venv (avoids PEP-668), tolerate missing pip, auto-add the client IP to the Postgres firewall, and never block infra provisioning on seeding failures. Re-runnable via 'azd hooks run postprovision'. - Set azure.yaml hooks.postprovision.continueOnError: true so a seeding hiccup doesn't fail an otherwise-successful 'azd up'. - Drop the dead 2-3 GB download branch (data ships with the repo, ~18 MB). - Fix .env.example: AZURE_OPENAI_EMBEDDING_DEPLOYMENT now matches the Bicep deployment (text-embedding-3-small, was text-embedding-ada-002). Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
agent/app.py shrinks from 609 -> 240 lines by:
- Moving Azure credential, ChatOpenAI, MultiServerMCPClient, MCP tool
fetch, and create_agent into a single Starlette lifespan. Each chat
request now reuses the same agent and token provider instead of
rebuilding all of them per request.
- Removing the local/production code duality. We now use
langchain-mcp-adapters in both modes; the OpenAI Responses-API native
MCP path ({"type": "mcp", ...}) is gone. One code path, one set of
semantics for stream events.
- Removing the image-generation sub-agent workaround for
langchain-ai/langchain#34136 (fixed in langchain v1). Image generation
is feature-flagged via ENABLE_IMAGE_GENERATION (default off) so the
default deploy is conservative.
- Dropping temperature=0.7 (gpt-5-mini is reasoning-style and ignores
or rejects non-default temperature).
- Adding bounded-retry/backoff for MCP startup (1,2,4,8,16s) so a
parallel start-up race in Container Apps doesn't crash-loop the
service unnecessarily.
- Reflecting readiness in /api/health (status, ready flag, mcp_tool_count).
agent/streaming.py (new, 214 lines) extracts the streaming response
parser into pure helpers driven by a layered chunk inspection:
tool_calls / additional_kwargs first, then content_blocks (LangChain v1
normalized format), then raw content. The state-machine for
{"status": ""} clear events is preserved exactly.
agent/test_streaming.py (new, 154 lines, 17 tests, pytest, 0.02s) covers
the NDJSON event contract with hand-built chunks for: text streaming,
reasoning blocks (ignored), tool_calls via tool_calls and via
additional_kwargs, server tool start/result, code_interpreter with
image data URLs and file outputs, response_metadata image outputs,
direct image blocks (data URL + raw base64), and JSON-stringified image
results from custom MCP tools.
agent/requirements.txt bumped to a coherent LC v1 set:
langchain >=1.0, langchain-openai >=1.0, langchain-core >=1.0,
langchain-mcp-adapters >=0.2.2 (latest published).
E2E verified locally against the deployed Azure MCP + OpenAI services:
agent boots in ~3s, loads 4 MCP tools, returns the documented NDJSON
event sequence (status -> empty status -> chunks -> done message), and
correctly executes a 'count products' query via the MCP tool against
the live Postgres (424 rows).
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
- Pin fastmcp to >=2.0,<3 (verified with 2.14.7); tools, lifespan, Context API, and http_app() all compatible. - Drop hand-rolled parse_postgres_url regex; pass DSN directly to asyncpg.create_pool. Add _safe_dsn() helper to URL-encode userinfo so passwords with '#' (auto-generated by azd) parse correctly. - Default embedding deployment to text-embedding-3-small (matches the Bicep that infra/main.bicep actually deploys). - Bump api_version default to 2024-12-01-preview; configurable via AZURE_OPENAI_API_VERSION env var. - Tighten SQL guard: also reject WITH-without-SELECT, MERGE, CALL, COPY; accept WITH-prefixed SELECTs (CTEs). - Add unit tests (mcp/test_app.py): 25 tests covering SQL validation + DSN encoding edge cases. - File trimmed 470 -> 311 lines (-34%); behaviour unchanged externally. Tested: - 25/25 unit tests pass - Docker build OK with fastmcp 2.14.7 - E2E against deployed Postgres (with '#' in pwd) + live MCP tools: 4 tools register, get_table_schemas returns 10 tables, execute_sales_query returns 424 products, DROP TABLE rejected.
- Replace marketing-style feature list with a 'What you'll learn' section - Add repository layout tree - Expand 'How it works' with real explanations of Responses API, MCP transport, Entra ID auth, and Bicep - Convert troubleshooting tips to a table; trim filler Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Pre-SQLite modernization changes that have been validated end-to-end on the Postgres flavor.
What's included
2dd8d08Speed up deploy: uv-based Dockerfiles, resilient postprovisionuvfor ~3× faster dependency installpostprovisionhook idempotent and resilient to partial failuresdf72730Modernize agent: lifespan init, unified MCP path, testable stream parserlifespanfor clean startup/shutdown of MCP client and agent8476cf9Phase 3: modernize MCP server (fastmcp 2.x, simpler DSN handling)fastmcp2.x APIsNot included
SQLite-backed minimal flavor and follow-up fixes — those will land in a separate PR once the chart-rendering issue is resolved.
Validation
azd upsucceeds end-to-end against Postgres; agent answers product, sales, and chart questions correctly.Co-authored-by: Copilot 223556219+Copilot@users.noreply.github.com