Meridian Knowledge & Access Agent
A controlled document corpus is ingested and split into section-aware chunks. Amazon Titan Text Embeddings v2 on Bedrock generates 1,024-dimensional embeddings; vectors and metadata are stored in PostgreSQL with pgvector. At query time, the system retrieves relevant passages using semantic similarity plus metadata signals, then generates an evidence-grounded response with conflict handling and an unsupported-answer fallback.
View on GitHubCorpus size
25 pages, 152 chunks
Evaluation cases
Clean/control, hard-conflict, near-conflict, and noise cases
Bedrock region
us-east-2
Automated tests
Test suite passing (count pending final verification)
Demo setup
Local Docker + Bedrock (preferred demo environment)
Possible production target
ECS Fargate / RDS — not continuously deployed
Retrieval
The corpus (25 pages, 152 chunks) is split into section-aware chunks so retrieved passages preserve their original document context rather than arbitrary fixed-length windows.
Amazon Titan Text Embeddings v2 (via Amazon Bedrock, us-east-2) generates 1,024-dimensional embeddings for each chunk. Vectors and their metadata are stored in PostgreSQL using pgvector, and queries are answered with semantic similarity search combined with metadata signals rather than similarity alone.
Conflict & freshness evaluation
The evaluation corpus was deliberately built to include clean/control cases, hard-conflict cases, near-conflict cases, and noise cases, so retrieval and response quality can be tested against ambiguous and contradictory source material — not only straightforward lookups.
When retrieved passages disagree, the system is designed to surface that conflict rather than silently picking one source.
Evidence grounding & fallback
Responses are generated from retrieved evidence rather than open-ended generation, with Llama 3.3 used for generation on top of the retrieved context.
When the corpus does not contain sufficient support for a question, the system is designed to fall back rather than produce an unsupported answer.
Limitations & roadmap
This is a personal portfolio project built against a fictional Meridian Financial Group document corpus — it is not Fidelity production work and should not be read as such.
Live Confluence integration, a custom MCP server, a Spring AI migration, and a workflow-automation agent are roadmap items only; they are not part of the current implementation.
A possible production target is ECS Fargate with RDS, but the project is not continuously deployed — local Docker plus Bedrock is the current demo environment.
Implemented
- Section-aware document chunking over a controlled document corpus
- 1,024-dimensional embeddings via Amazon Titan Text Embeddings v2 on Bedrock
- Vector + metadata storage and retrieval in PostgreSQL with pgvector
- Evidence-grounded response generation with conflict detection
- Fallback behavior when the corpus does not support an answer
Roadmap (planned, not implemented)
- Live Confluence integration
- Custom MCP server
- Spring AI migration
- Workflow-automation agent