Thread existing JSON schemas from prompt definitions through the
text-completion service to LLM backends' native structured output
APIs. When a prompt has response-type "json" and a strict-mode
compatible schema, the LLM constrains token selection at the logit
level to guarantee schema-valid output.
Wire-level changes:
- Add response_format and schema fields to TextCompletionRequest
- Update translator to encode/decode new fields
- Pass new fields through LlmService, TextCompletionClient, and
PromptManager
Runtime schema compatibility checker:
- New is_strict_mode_compatible() utility validates schemas against
LLM provider constraints (additionalProperties, required fields,
no unsupported constraints, no open-ended objects)
- Per-prompt eligibility decision: compliant schemas use structured
output, non-compliant schemas fall back to free-text + post-hoc
validation
LLM backend implementations:
- OpenAI: response_format with json_schema, variant-aware top-level
array rejection (openai variant blocks, llama/vllm variants allow)
- New vllm variant for the OpenAI backend
- vLLM (dedicated): response_format in raw HTTP body
- Ollama: format=<schema> parameter
- Claude: tool-use trick (forced tool call with schema as input_schema)
- Mistral: native json_schema response_format
- Llamafile, LM Studio: OpenAI SDK response_format
- Azure OpenAI: AzureOpenAI SDK response_format
- Azure serverless: response_format in raw HTTP body
- TGI: response_format in raw HTTP body
- VertexAI Gemini: response_mime_type + response_schema
- VertexAI Claude: tool-use trick
- Google AI Studio: response_mime_type + response_schema
- Bedrock, Cohere: signature-only (no structured output yet)
Post-hoc jsonschema.validate() retained as defence-in-depth.
Tech spec added: docs/tech-specs/structured-output.md
Update tests
* Changed schema for Value -> Term, majorly breaking change
* Following the schema change, Value -> Term into all processing
* Updated Cassandra for g, p, s, o index patterns (7 indexes)
* Reviewed and updated all tests
* Neo4j, Memgraph and FalkorDB remain broken, will look at once settled down
* Tidy up duplicate tech specs in doc directory
* Streaming LLM text-completion service tech spec.
* text-completion and prompt interfaces
* streaming change applied to all LLMs, so far tested with VertexAI
* Skip Pinecone unit tests, upstream module issue is affecting things, tests are passing again
* Added agent streaming, not working and has broken tests
- Keeps processing in different flows separate so that data can go to different stores / collections etc.
- Potentially supports different processing flows
- Tidies the processing API with common base-classes for e.g. LLMs, and automatic configuration of 'clients' to use the right queue names in a flow
* Break out enums for different model types
* Add model detection for inference profiles in US and EU
* Encapsulate model handling, make it easier to manage
* - More AWS Boto3 settings (profile and session key)
- Align environment variable and profile setting names with AWS
conventions.
Hopefully this should be able to run from an EC2 instance just with role
setting.
* Tweak naming to all make sense, added rate limit detect
* - Refactored retry for rate limits into the base class
- ConsumerProducer is derived from Consumer to simplify code
- Added rate_limit_count metrics for rate limit events
* Add rate limit events to VertexAI and Google AI Studio
* Added Grafana rate limit dashboard
* Add rate limit handling to all LLMs
- Change templates to interpolate environment variables in docker compose
- Change templates to invoke secrets for environment variable credentials in K8s configuration
- Update LLMs to pull in credentials from environment variables if not specified
